Charging cabinet operation state monitoring method, device and system based on internet of things data

Abstract

The application provides a charging cabinet operation state monitoring method, device and system based on Internet of Things data, relates to the technical field of artificial intelligence, and performs data normalization based on an absolute safety boundary in combination with local statistical characteristics, can not only quickly filter extreme abnormal data, but also can adapt to the data characteristics of each channel, maximizes the retention of core time sequence characteristics while eliminating dimensional differences. A channel interaction fusion vector is further constructed, and a boundary violation sensitive feature is extracted to perform state monitoring. The channel interaction fusion vector can capture the corresponding composite abnormal mode of the data and improve the comprehensiveness of abnormal identification; the boundary violation sensitive feature can accurately capture early signs of failure, realizing the change from passive response to active early warning. In summary, the application can break through the identification limitations of the prior art, accurately identify the operation state of the charging cabinet from massive high-dimensional time sequence data, and can strengthen the early warning ability of operation safety, improve the intelligentization and scale level of operation and maintenance management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a charging cabinet operation state monitoring method, device and system based on Internet of Things data. BACKGROUND

[0002] With the rapid popularization of electric vehicles, electric bicycles and various portable electronic devices, charging cabinets as centralized and shared electric energy supply infrastructure have increased in deployment scale. The operation safety and reliability of charging cabinets are directly related to user property, power safety and even public safety. Once abnormal states such as overvoltage and overcurrent occur, it may cause equipment damage, fire and other serious consequences. Therefore, real-time and accurate online monitoring and early fault warning of the operation state of the charging cabinet have become an urgent and key technical requirement in operation and maintenance management.

[0003] Currently, the monitoring technology based on Internet of Things provides a data basis for this demand. By deploying a sensor network inside the charging cabinet, time series data of key physical quantities such as current, voltage and power during the charging process can be collected in real time. However, how to automatically and accurately identify various normal and abnormal operation states from these massive and high-dimensional time series data still faces a series of technical challenges. SUMMARY

[0004] The purpose of the present application is to provide a charging cabinet operation state monitoring method, device and system based on Internet of Things data, which can break through the identification limitations of the prior art and automatically and accurately identify various normal and abnormal operation states from massive and high-dimensional time series data.

[0005] In a first aspect, the present application provides a charging cabinet operation state monitoring method based on Internet of Things data, which comprises: determining the operation monitoring data of a target charging cabinet through the Internet of Things sensor network of the target charging cabinet; based on a preset absolute safety boundary and the local statistical features corresponding to each monitoring sequence of each monitoring channel in the operation monitoring data, performing normalization processing on the operation monitoring data to generate target monitoring data; constructing a channel interaction fusion vector of the target monitoring data and determining a boundary violation sensitive feature of the target monitoring data; based on the channel interaction fusion vector and the boundary violation sensitive feature, performing classification prediction on the target monitoring data to determine the operation state classification result of the target charging cabinet.

[0006] With reference to the first aspect, in a first implementation form of the first aspect, the step of generating the target monitoring data based on the preset absolute safety boundary and the local statistical features corresponding to the monitoring sequences of each monitoring channel in the operation monitoring data, comprises: determining the local average level corresponding to the monitoring sequences of each monitoring channel in the operation monitoring data based on the preset local window length; determining the dynamic normalization boundary corresponding to the local average level based on the preset absolute safety boundary; and performing normalization processing on the monitoring sequences of each monitoring channel based on the dynamic normalization boundary to construct the target monitoring data corresponding to the operation monitoring data.

[0007] With reference to the first aspect, in a second implementation form of the first aspect, the step of constructing the channel interaction fusion vector of the target monitoring data, comprises: performing inter-channel information fusion on the monitoring sequences of each monitoring channel of the target monitoring data by using the preset coupling kernel tensor to generate the intermediate tensor corresponding to the target monitoring data; performing singular value decomposition on the intermediate tensor to determine the first change mode of the target monitoring data with respect to the time joint dimension and the second change mode of the target monitoring data with respect to the data depth dimension; and performing data filtering processing on the target monitoring data based on the first singular vector matrix corresponding to the first change mode and the second singular vector matrix corresponding to the second change mode to generate the channel interaction fusion vector of the target monitoring data.

[0008] With reference to the first aspect, in a third implementation form of the first aspect, the step of determining the boundary violation sensitive feature of the target monitoring data, comprises: determining the out-of-bound event data corresponding to the target monitoring data based on the preset out-of-bound parameter; determining the boundary proximity data corresponding to the target monitoring data based on the normalized relative boundary corresponding to the normalization processing; and determining the boundary violation sensitive feature of the target monitoring data based on the out-of-bound event data and the boundary proximity data.

[0009] With reference to the first aspect, in a fourth implementation form of the first aspect, the step of performing classification prediction on the target monitoring data based on the channel interaction fusion vector and the boundary violation sensitive feature to determine the operation state classification result of the target charging cabinet, comprises: fusing the boundary violation sensitive feature into the channel interaction fusion vector to construct the focused feature tensor corresponding to the target monitoring data; and performing classification prediction on the focused feature tensor by using the pre-constructed operation state monitoring model to determine the operation state classification result corresponding to the target charging cabinet.

[0010] With reference to the first aspect, the embodiment of the present application provides a fifth implementation manner of the first aspect, and the step of fusing the boundary violation sensitive feature into the channel interaction fusion vector to construct the focused feature tensor of the target monitoring data comprises: fusing the boundary violation sensitive feature into the channel interaction fusion vector to construct an initial fusion feature tensor by using a preset gating fusion mechanism; and constructing the focused feature tensor by adaptively weighting the initial fusion feature tensor based on a preset multi-dimensional attention weight.

[0011] With reference to the first aspect, the embodiment of the present application provides a sixth implementation manner of the first aspect, and the step of classifying and predicting the focused feature tensor by using a pre-constructed operation state monitoring model to determine the operation state classification result corresponding to the target charging cabinet comprises: determining a probability distribution prediction result of each monitoring channel of the focused feature tensor based on a corresponding mode summary vector by using a multi-prototype metric learning mechanism; and determining the operation state classification result corresponding to the target charging cabinet based on the probability distribution prediction result.

[0012] With reference to the first aspect, the embodiment of the present application provides a seventh implementation manner of the first aspect, and the operation state monitoring model is constructed based on the channel interaction fusion vector and the boundary violation sensitive feature of a preset training sample; the loss function of the operation state monitoring model comprises a violation perception weighted classification loss and a coupled mode filtering consistency regularization loss; wherein the violation perception weighted classification loss is calculated based on a dynamic weight of the training sample, and the dynamic weight is determined based on the boundary violation sensitive feature; and the coupled mode filtering consistency regularization loss is used for consistency constraint on a filtering parameter corresponding to the channel interaction fusion vector.

[0013] In the second aspect, the embodiment of the present application provides a charging cabinet operation state monitoring device based on Internet of Things data, and the device comprises: a data monitoring module, configured to determine operation monitoring data of a target charging cabinet through an Internet of Things sensor network of the target charging cabinet; a data preprocessing module, configured to perform normalization processing on the operation monitoring data based on a preset absolute safety boundary and a local statistical feature corresponding to a monitoring sequence of each monitoring channel of the operation monitoring data, to generate target monitoring data; a data processing module, configured to construct a channel interaction fusion vector of the target monitoring data and determine a boundary violation sensitive feature of the target monitoring data; and an execution module, configured to classify and predict the target monitoring data based on the channel interaction fusion vector and the boundary violation sensitive feature, to determine an operation state classification result of the target charging cabinet.

[0014] In the third aspect, the embodiment of the present application provides a charging cabinet operation state monitoring system based on Internet of Things data, which is configured with the device of the above-mentioned embodiment and is used for executing the method of any one of the above-mentioned embodiments.

[0015] The embodiment of the present application brings the following beneficial effects: the embodiment of the present application provides a charging cabinet operation state monitoring method, device and system based on Internet of Things data, which can quickly filter extreme abnormal data through an absolute safety boundary, avoid the interference of such data from the source, and adapt to the data characteristics of each channel through local statistical characteristics, while eliminating the dimensional difference and maximizing the retention of core timing characteristics. For this purpose, a channel interaction fusion vector is constructed, which can capture complex abnormal patterns that cannot be reflected by a single channel, and can improve the comprehensiveness of abnormal identification; the boundary violation sensitive feature can break the limitations of the post-determination of "exceeding the standard and alarming", accurately capture early warning signals of faults, and reserve sufficient time for operation and maintenance disposal, from passive response to active early warning. In summary, the embodiment of the present application can effectively solve the technical challenges of automatic and accurate identification of the operation state of the charging cabinet under the condition of massive high-dimensional time series data in the prior art, which can not only strengthen the early warning capability of the operation safety of the charging cabinet, but also improve the intelligentization and scale level of operation and maintenance management.

[0016] Other features and advantages of the present application will be described in the following description, or can be inferred from the description or determined without doubt, or can be known by implementing the above-mentioned technologies of the present application.

[0017] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creating any creative labor.

[0019] Figure 1 A flowchart of a charging cabinet operation state monitoring method based on Internet of Things data provided by the embodiment of the present application is provided.

[0020] Figure 2 A flowchart of another charging cabinet operation state monitoring method based on Internet of Things data provided by the embodiment of the present application is provided.

[0021] Figure 3 A working principle of a boundary violation sensitive feature and a response diagram of the boundary violation sensitive feature to an abnormal state provided by the embodiment of the present application are provided.

[0022] Figure 4 A classification prediction flowchart provided by the embodiment of the present application is provided.

[0023] Figure 5A structural schematic diagram of a charging cabinet operation state monitoring device based on Internet of Things data is provided for an embodiment of the present application.

[0024] Figure 6 A structural schematic diagram of an electronic device is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.

[0026] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0027] The embodiments of the present application provide a charging cabinet operation state monitoring method, device and system based on Internet of Things data, which can break through the identification limitations of the prior art and automatically and accurately identify various normal and abnormal operation states from massive, high-dimensional time series data.

[0028] For ease of understanding, first, a charging cabinet operation state monitoring method based on Internet of Things data provided by an embodiment of the present application will be described with reference to Figure 1 The method comprises the following steps:

[0029] In step S102, operation monitoring data of a target charging cabinet is determined through an Internet of Things sensor network of the target charging cabinet.

[0030] The target charging cabinet is a specific monitoring subject in the embodiments of the present application that needs to carry out real-time monitoring of operation state and early warning of faults, which can be used as an independent individual to realize fine monitoring for a single charging cabinet, or as a basic monitoring unit in a batch monitoring scene. In the key working condition nodes inside the target charging cabinet, including but not limited to core positions such as charging modules, circuit interfaces and heat dissipation systems, an Internet of Things sensor network is pre-deployed. The network, as a data collection core carrier, relies on Internet of Things communication technology to collect various state data in the operation process of the target charging cabinet in real time, and transmits the collected original data to a back-end processing unit in real time. The Internet of Things sensor network is a distributed sensing network, which can be composed of multiple different types of sensors to collect multi-dimensional physical quantity data such as current, voltage, power, temperature and humidity that are strongly related to the operation safety of the charging cabinet.

[0031] The corresponding operation monitoring data is the original time series data collected by the above-mentioned Internet of Things sensor network, to accurately reflect the real-time state of the whole process of the operation of the target charging cabinet. Among them, it can include current, voltage, power and other multi-dimensional physical quantity data closely related to the safe operation of the charging cabinet, and the overall presents high-dimensional and time series characteristics.

[0032] In step S104, based on the preset absolute safety boundary and the local statistical features corresponding to each monitoring channel of the operation monitoring data respectively, the operation monitoring data is normalized to generate target monitoring data. The preset absolute safety boundary refers to the safety threshold range of each monitoring physical quantity preset based on the charging cabinet equipment parameters, industry safety standards and engineering practical experience, which is used as a bottom constraint to judge whether the data is in an extremely dangerous state. The data stream exceeding the boundary can be directly judged as a serious anomaly and processed first. The data acquisition link of a single sensor or a group of sensors of the same type in the above-mentioned sensor network is a monitoring channel, and its monitoring sequence is a time series data set collected by a single monitoring channel in a continuous time dimension, reflecting the dynamic change trend of the corresponding physical quantity during the charging process (for example, the sequence of current values collected every second within 1 minute of a certain current channel). The local statistical features are statistical indicators calculated within a preset time window (such as 10 seconds, 30 seconds) for the monitoring sequence of a single monitoring channel, and the core includes mean, variance, peak, valley, trend slope, etc., which can reflect the local fluctuation rule and distribution characteristics of the time series data, avoiding the accidental interference of single time point data.

[0034] In summary, the embodiment of the present application can quickly identify and filter extreme abnormal data exceeding the safety threshold based on the preset absolute safety boundary as a bottom rigid constraint, avoiding the interference of extreme data on subsequent feature extraction, classification and prediction from the source; at the same time, the differential processing is carried out based on the local statistical features of each monitoring channel, rather than using a unified standardization rule, so that the normalization process can accurately adapt to the unique fluctuation rule and distribution characteristics of different physical quantity time series data, and the core time series features related to the charging cabinet operation state in the original data are retained to the maximum extent. In summary, the dimensional difference and distribution unevenness of data from different monitoring channels can be eliminated, the high-dimensional heterogeneous time series data can be mapped to a unified numerical interval, and the accidental interference of single time point data can be filtered, so that the generated target monitoring data is more suitable for the actual operation condition of the charging cabinet.

[0035] In step S106, a channel interaction fusion vector of the target monitoring data is constructed, and a boundary violation sensitive feature of the target monitoring data is determined.

[0036] The channel interaction fusion vector is a high-dimensional feature vector formed after the target monitoring data of multiple monitoring channels are cross-dimensionally associated and integrated (for example, constructed by a feature fusion algorithm, in an embodiment, an attention mechanism, convolution fusion, matrix splicing, etc. can be adopted), which can capture the collaborative change relationship between different channel data (for example, the association of abnormal fluctuations in the current channel and mutations in the voltage channel).

[0037] The boundary violation sensitive feature is a critical state feature in the target monitoring data that approaches or slightly touches the preset absolute safety boundary, such as including the fluctuation amplitude, duration, rate of change of critical data, and the collaborative occurrence mode of multi-channel critical state, etc., to capture early warning signals before failure occurs, and avoid the limitations of post-judgment of "exceeding the standard to alarm".

[0038] In step S108, the target monitoring data is classified and predicted based on the channel interaction fusion vector and the boundary violation sensitive feature, and the running state classification result of the target charging cabinet is determined.

[0039] Based on the channel interaction fusion vector, the collaborative correlation rules captured can be fully utilized to accurately identify complex abnormal patterns that cannot be reflected by single channel features, avoid missed detection problems caused by traditional single parameter monitoring, and improve the comprehensiveness of abnormal state recognition; based on the boundary violation sensitive feature, the critical abnormal signal close to the absolute safety boundary can be accurately anchored, the early warning of failure can be effectively captured, and the limitations of "exceeding the standard to alarm" are broken, which reserves sufficient time for operation and maintenance disposal. In summary, based on the complementary information of the two types of features, the running state corresponding to the target monitoring data can be accurately classified, which can not only clearly distinguish between normal operation, slight abnormality, severe abnormality and other basic states, but also further locate specific abnormal types such as overvoltage and overcurrent, thereby improving the accuracy and directionality of state judgment. At the same time, the whole process of the embodiment of the present application is automatically executed, which can efficiently process batch monitoring data of charging cabinets, can adapt to large-scale cluster operation and maintenance scenarios, greatly reduces the cost of manual operation and maintenance, and improves the operation and maintenance efficiency.

[0040] In summary, the embodiment of the present application can quickly filter extreme abnormal data through the absolute safety boundary, avoid the interference of such data from the source, and adapt to the data characteristics of each channel through local statistical characteristics, while eliminating the dimensional differences and maximizing the retention of core timing characteristics. In this regard, the construction of a channel interaction fusion vector can capture complex abnormal patterns that cannot be reflected by a single channel, thereby improving the comprehensiveness of abnormal identification. The boundary violation sensitive features can break the limitations of post-determination of "exceeding the standard", accurately capture early signs of failure, and provide sufficient time for operation and maintenance, thereby changing from passive response to active early warning. In summary, the embodiment of the present application can effectively solve the technical challenges of automatic and accurate identification of the running state of the charging cabinet in the prior art under the condition of massive high-dimensional time series data, which can not only strengthen the early warning capability of the running safety of the charging cabinet, but also improve the intelligentization and scale level of operation and maintenance management.

[0041] On the basis of the above-mentioned embodiment, the embodiment of the present application further provides another charging cabinet running state monitoring method based on Internet of Things data, referring to Figure 2 The method comprises the following steps:

[0042] In step S202, the running monitoring data of the target charging cabinet is determined through the Internet of Things sensor network of the target charging cabinet.

[0043] On the basis of the above-mentioned embodiment, the corresponding data can be synchronously collected by the Internet of Things sensor network deployed in the charging cabinet at a fixed sampling frequency. In the embodiment of the present application, it mainly includes current sequence, voltage sequence and power sequence. The current, voltage and power sequence are three types of core data in the charging cabinet running state monitoring, which can directly reflect the electrical safety and energy interaction nature of the charging process. The current sequence can represent the load change and overcurrent risk, the voltage sequence can indicate the power supply stability and insulation abnormality, and the power sequence can comprehensively reflect the energy conversion efficiency and abnormal power consumption mode. The three types of data together constitute a multi-dimensional time series feature basis. Further, by determining the coupling relationship, the complex abnormal pattern can be captured, and by combining the absolute safety boundary and the local statistical characteristics, the early failure precursor can be accurately identified and the cooperative anomaly can be efficiently distinguished, thereby significantly improving the comprehensiveness and early warning capability of the state monitoring.

[0044] Among them, the monitoring data of each charging process constitutes an original data sample, which contains the complete trajectory of the three physical quantities changing with time from the beginning to the end of charging or within a certain monitoring period.

[0045] Correspondingly, the original data is defined as multi-channel time series data, and the original data sample is set as , for the three synchronous monitoring sequences of current, voltage and power, the corresponding dimension is , which are synchronously collected by the sensors of the charging cabinet. Each row represents a sequence of a physical quantity changing over time. represents the length of the time sequence, which is the total length of the time sequence of a single sample; represents the 0th row of the original data sample , corresponding to the entire current monitoring time sequence. represents the 1st row of the original data sample , corresponding to the entire voltage monitoring time sequence. represents the 2nd row of the original data sample , corresponding to the entire power monitoring time sequence.

[0046] Further, the corresponding data is preprocessed (normalized). The current, voltage, and power monitoring sequence data of the charging cabinet have clear physical safe operation upper and lower boundary limits. The actual running data fluctuates within the normal range, while the abnormal state often shows slight out-of-bound or sharp mutation near the boundary. The conventional global maximum and minimum normalization blurs the physical boundary meaning and is too sensitive to the abnormal points in the sequence, resulting in excessive compression of the normal fluctuation after normalization. The Z-score standardization cannot reflect the absolute safe range of the data. These deficiencies weaken the model's ability to identify boundary violations and abnormal fluctuations.

[0047] For the above step S104, the application adopts a dynamic double-boundary normalization method. Not only the prior physical boundary (i.e., the absolute safe boundary) is referred to, but also the normalization boundary at each time is dynamically calculated in combination with the local statistical characteristics. Further, by extending the normalization value range beyond the reference interval, the out-of-bound information is retained and even highlighted, thereby enhancing the model's sensitivity to behaviors near the boundary and abnormal states. In specific implementation, the following steps S204-S208 are referred to.

[0048] Step S204, based on the preset local window length, determines the local average level corresponding to each monitoring channel of the running monitoring data.

[0049] The local average level can represent the running data reference of the corresponding monitoring channel at different time periods, reflecting the local stationary characteristics of the time sequence data. The corresponding local average level can be calculated at each time through a sliding time window mechanism, for example, the preset local window length is used to define the statistical range of the local data. Its value can be adaptively adjusted according to the running conditions of the charging cabinet and the monitoring accuracy requirements. In one embodiment, the value example is 11, which ensures that enough data samples are covered to reflect the local characteristics, and avoids the problem of loss of time sequence details caused by too long window.

[0050] For the monitoring sequence corresponding to the c-th monitoring channel, each time t can be traversed along the time-series dimension. Taking time t as the center of the window, the sequence of that channel with length L centered at t is extracted. w A local data window, i.e., a subsequence containing data points at time t and their immediate vicinity, is used to accurately extract local data near time t. Furthermore, the arithmetic mean of all data points within the extracted local data window can be calculated to obtain the local average level of the c-th channel sequence at time t, denoted as [the average level]. , indicating the first The channel sequence with the first Centered on time, with a local window length of The arithmetic mean of the data within the window is used. This arithmetic mean can effectively suppress the interference of random fluctuations of a single data point, accurately depicting the local stationary benchmark and central tendency of the data in the c-th channel near time t. Furthermore, the above local data window can be slid along the time-series dimension according to a preset step size, and the above truncation and mean calculation operations can be repeated to generate the sequence corresponding to the c-th channel at each time step. This ultimately forms a local average level sequence with the same length as the original monitoring sequence.

[0051] Step S206: Based on the preset absolute safety boundary, determine the dynamic normalization boundary corresponding to the local average level.

[0052] Furthermore, for each monitoring channel and each moment, the upper and lower boundaries of dynamic normalization are calculated. These boundaries are determined by a predefined absolute safety hard boundary and the local statistical characteristics of the data at that moment. The preset absolute safety boundary is the safety threshold range of the physical quantities of each monitoring channel (such as upper / lower current limit, upper / lower voltage limit), which is the rigid bottom line to ensure the safe operation of the charging cabinet. Combining the aforementioned local average level, the normalization boundary that dynamically adjusts with the local average level can be calculated. For example, when the local average level is too high, the dynamic normalization boundary moves closer to the upper limit of the absolute safety boundary; when the local average level is too low, the dynamic normalization boundary moves closer to the lower limit of the absolute safety boundary. Based on this, the dynamic normalization boundary can both anchor the safety bottom line of the absolute safety boundary, ensuring that the normalization process respects physical constraints, and adapt to the local fluctuation characteristics of the data of each channel, achieving a balance between safety constraints and accurate data mapping.

[0053] In one implementation, it can be calculated using the following formula:

[0054]

[0055]

[0056] In the formula, Indicates the first The absolute safety lower limit of the first channel is defined in advance according to the charging cabinet specifications. The absolute safety lower limit of each channel is a constant hard boundary that does not change with time. It is used to constrain the dynamic boundary to not be lower than this value. For example, the current channel can be set to 0 amps. Indicates the first The absolute safety limit for each channel is defined in advance according to the charging cabinet specifications. The absolute safety upper limit of each channel, as a constant hard boundary that does not change over time, is used to constrain dynamic boundaries from exceeding this value. For example, a voltage channel could be 300 volts. Indicates the channel index. Represents the current path. Represents the voltage channel. Represents the power channel; Indicates a point-in-time index. . Indicates the first The first channel in the The dynamic normalized lower boundary at time; Indicates the first The first channel in the The dynamic normalized upper bound of time. This represents the boundary scaling factor, a hyperparameter greater than 0, used to control the dynamic boundary bandwidth composed of local statistics. An example value is 2.0. This indicates the operation of retrieving the maximum value; This indicates the operation of taking the minimum value. Indicates the first The channel sequence with the first Centered on time, with a local window length of The standard deviation of the data within the window is used to characterize the local fluctuation amplitude of the data near that moment. The local window length is... An example value for is 11. Furthermore, the boundary can be extended using a symmetrically padded sequence to ensure that the window length is constant at each time step. This avoids statistical distortion.

[0057] Step S208: Based on the dynamic normalization boundary, the monitoring sequence of each monitoring channel is normalized to construct the target monitoring data corresponding to the running monitoring data.

[0058] The original data samples are normalized using the calculated dynamic normalization upper and lower boundaries, mapping their feature values ​​to the vicinity of the baseline interval. Notably, the normalized values ​​are allowed to exceed the range without strict constraints on the output value range. The range is defined, thus retaining out-of-bounds information as an anomaly signal, represented as:

[0059]

[0060] wherein, denotes normalized data The feature value of the i-th channel at the j-th time point is converted into a relative value based on the dynamic boundary by dynamic double boundary normalization mapping. denotes the value of the i-th channel at the j-th time point; denotes a minimum positive number, used to prevent the denominator from being zero and ensure numerical stability, and an example of the value is denotes a minimum positive number, used to prevent the denominator from being zero and ensure numerical stability, and an example of the value is It should be noted that when is located within the dynamic boundary, the value of is approximately in the interval If is lower than the lower boundary of dynamic normalization , If is higher than the upper boundary of dynamic normalization , The out-of-bound information retention mechanism can enhance the model's ability to identify abnormal behavior.

[0061] Further, the working principle of the above boundary violation sensitive feature and its response to abnormal state are schematically illustrated, referring to Figure 3 ​​​​The embodiments of the present application show the corresponding working principle and the response to the abnormal state through the combination of three subgraphs to visualize the process of calculating the boundary violation sensitive feature vector in the embodiments of the present application, and prove that the feature can effectively quantify the abnormal risk and highlight the violation behavior near the boundary. Among them, the experimental analysis includes two scenarios: normal charging current time sequence and charging current time sequence containing an overcurrent anomaly. The safe lower boundary of the current is defined as 5 amperes and the safe upper boundary is defined as 25 amperes in the graph, which is indicated by a red dashed line. The first subgraph shows the fluctuation of the normal state current within the safe boundary, and the green filled area represents the safe operation interval. The second subgraph shows an overcurrent anomaly case, in which the current value exceeds the safe upper boundary during the period of 8 seconds to 12 seconds, and the red filled area in the graph clearly indicates the "abnormal boundary crossing area". The third subgraph compares and calculates the boundary violation sensitive feature value corresponding to each time under the two states. The experimental results show that in the normal state of the first subgraph, the boundary violation sensitive feature value curve (blue) always maintains a low level and fluctuates gently. While in the abnormal state of the second subgraph, the corresponding feature value curve (red) is close to the blue curve before 8 seconds and after 12 seconds, but in the abnormal boundary crossing area (8 seconds to 12 seconds), the red curve rises sharply, forming a significant peak, indicating that when the current exceeds the absolute safe boundary, the boundary violation sensitive feature value calculated by the embodiments of the present application will produce a sharp positive response. More importantly, this feature is not only sensitive to explicit boundary crossing, but also quantifies the degree of data points approaching the boundary, providing an early warning for potential risks.

[0062] In step S210, a channel interaction fusion vector of the target monitoring data is constructed, and a boundary violation sensitive feature of the target monitoring data is determined.

[0063] Although the current, voltage and power time sequence data of the normalized data have retained the physical boundary information, the three channels are still independent, the running state of the charging cabinet is essentially determined by the coupling dynamics of multiple physical quantities, and the conventional feature fusion method uses a simple channel splicing method, which cannot mine deep temporal and spatial coupling correlations between channels, resulting in difficulty for the model to capture joint feature patterns that determine the running state.

[0064] To this end, the present application constructs a corresponding channel interaction fusion vector. Among them, the embodiments of the present application adopt a coupled tensor enhancement method, introduce a learnable coupling kernel tensor to model and enhance the interaction information between channels, and generate a feature tensor (i.e. intermediate tensor) rich in coupling relationship. Further, through adaptive singular value filtering, noise is removed and focused, and robust and significant joint feature patterns are extracted. Correspondingly, the embodiments of the present application construct the channel interaction fusion vector through the following steps:

[0065] 1) Use the preset coupling kernel tensor to perform inter-channel information fusion on the monitoring sequence of each monitoring channel of the target monitoring data to generate the intermediate tensor corresponding to the target monitoring data.

[0066] The normalized data mentioned above is a 3-channel single-depth feature. In this embodiment of the invention, it is further transformed into a 3-channel feature. Deep feature enhancement aims to specifically mine feature information at different levels, enriching the dimensions of feature representation. For example, it can enhance features with dimensions of... Normalized data Reconstructing into dimensions is A third-order tensor, with the last dimension initialized to 1; simultaneously, the augmented depth dimension is set to... , represents the number of coupling patterns to be mined, is a preset hyperparameter, and its value determines the number of potential coupling feature patterns mined and generated from the original three-channel data. An example value is 32.

[0067] Furthermore, by introducing a trainable coupled kernel tensor, the original 3-channel single-depth features can be transformed into 3-channel D-depth enhanced features through tensor multiplication operations. This allows information from each channel to interact and fuse with other channels at multiple depths, generating an intermediate tensor, represented as:

[0068]

[0069] In the formula, The intermediate tensor, through learnable coupling kernels that fuse inter-channel information, represents the joint representation of multi-channel time-series data across multiple depth modes, with a dimension of [missing information]. . Represents intermediate tensor In the Time, Number The vector at depth is a 3-dimensional column vector, representing the coupled feature vectors of the three channels at a specific time and depth, i.e., the first... Joint representation of various coupling modes. The coupled kernel tensor is a trainable parameter with dimension O(n). The coupling relationship between channels can be modeled using tensor multiplication. Represents the coupled kernel tensor The middle corresponds to the input channel and deep mode A 3D column vector representing the original first... The enhanced three channels are all in the first... Contribution weights in deep mode. Indicates the first The bias vector corresponding to the depth is a trainable parameter. Among them, Corresponding to the third-order tensor of the above reconstruction .

[0070] 2) Perform singular value decomposition on the intermediate tensor to determine a first variation mode of the target monitoring data for the time joint dimension and a second variation mode of the target monitoring data for the data depth dimension.

[0071] 3) Perform data filtering processing on the target monitoring data based on a first singular vector matrix corresponding to the first variation mode and a second singular vector matrix corresponding to the second variation mode, to generate a channel interaction fusion vector of the target monitoring data.

[0072] The singular value decomposition processing can purify key feature modes from the high-dimensional coupled intermediate tensor, suppress noise interference, and realize accurate layering and quantitative representation of different levels of feature information. The intermediate tensor fuses three-dimensional information of the charging cabinet monitoring data, including the channel dimension (3 physical quantities), the time dimension (T time points), and the depth dimension (D layers of enhanced features). Different dimensional data are mutually coupled, direct analysis is easily disturbed by noise, and it is difficult to locate the core mode. In order to extract the most significant coupled mode from the intermediate tensor and suppress noise, the intermediate tensor is first reshaped into a two-dimensional matrix to map high-dimensional coupled information to a unified two-dimensional space. Then, economic singular value decomposition is performed to decompose the matrix into a left singular vector matrix, a singular value matrix, and a right singular vector matrix, realizing decoupling of the cross-dimensional coupled mode. Among them, represents the left singular vector matrix, representing the main variation mode of the data in the channel-time joint dimension (i.e., the first variation mode, such as the coordinated fluctuation mode of current-voltage over time). The column vectors constitute an orthogonal basis, and the dimension is ; represents the right singular vector matrix, and the column vectors span a space representing the main variation mode of the data in the depth dimension (i.e., the second variation mode, such as the weight distribution mode of different depth enhanced features). The column vectors constitute an orthogonal basis, and the dimension is .

[0073] Further, represents the singular value matrix, which can be quantified and filtered based on the singular value to suppress noise redundancy. The matrix is a diagonal matrix of , i.e. , whose diagonal elements are singular values, and satisfy The size of the singular value represents the importance of the corresponding mode: the first few larger singular values correspond to the most significant core feature mode in the data, and the last few smaller singular values correspond to noise or redundant information. represents the th singular value, which is a positive value representing the importance or energy contribution of the mode represented by the corresponding singular vector in the data.

[0074] Among the above parameters, denotes the rank of the following determined two-dimensional matrix The economic singular value decomposition can filter noise interference, realize data dimension reduction, accurately retain the core feature patterns of the channel-time and depth two levels, and avoid interference of irrelevant information on subsequent state recognition by setting the rank r (much smaller than the dimension of the original matrix) and retaining only the first r key singular values and the corresponding singular vectors. Correspondingly, the two-dimensional matrix can be represented as: . In the formula, denotes the intermediate two-dimensional matrix, which is obtained by reshaping the intermediate tensor , and the dimension is The reshaping operation combines the channel and time dimensions into rows and the depth dimension as columns; denotes the left singular vector matrix, denotes the right singular vector matrix. denotes the transpose of .

[0075] Further, the singular values obtained by decomposition can be subjected to soft threshold filtering to attenuate small singular values that may correspond to noise, while retaining large singular values that correspond to main coupling modes, which is represented as:

[0076]

[0077] In the formula, denotes the first singular value after soft threshold filtering, which is used to construct a new singular value matrix, so as to attenuate noise patterns and retain main patterns when reconstructing the filtering matrix. is a sign function, and since , ; denotes an adaptive soft threshold value used to control the shrinkage amount in soft threshold filtering, and the calculation method is represented as . denotes a proportional coefficient, which is a hyperparameter for controlling the filtering strength, and an example value is 0.6. In summary, this process can attenuate small singular values by retaining the sign of the singular value and applying a soft threshold function in combination with a maximum value comparison.

[0078] Further, the singular values after soft threshold filtering can be used to form a new diagonal matrix, which is operated with the original left and right singular vector matrices to obtain a denoised and focused reconstructed filtering matrix, which is represented as: . In the formula, denotes the reconstructed filtering matrix, which realizes filtering of singular values corresponding to noise and highlights main coupling modes, so as to obtain a denoised and focused feature representation, and the dimension is . denotes a diagonal matrix with singular values after soft threshold filtering as diagonal elements, that is .

[0079] Further, the reconstructed filter matrix can be reshaped into a three-dimensional format to obtain an enhanced feature tensor (that is, a channel interaction fusion vector), with a dimension of . The enhanced feature tensor adds information in two dimensions of cross-channel time sequence correlation features and depth level weight features to the original normalized features, so that the feature representation is upgraded from single-dimensional time sequence data to multi-dimensional, hierarchical, and weighted structured features, significantly enriching the feature representation dimension and improving the feature discrimination between normal and abnormal states (especially complex abnormal and early critical abnormal states). In summary, the embodiment of the application can construct a channel interaction fusion vector that not only retains the time sequence information of the original channels, but also embeds significant inter-channel coupling patterns after adaptive filtering.

[0080] The running state of the charging cabinet depends not only on instantaneous features, but also on the time evolution trend and the topology pattern formed by the combination of multiple features. Conventional classification methods usually flatten the features and directly input them into a fully connected network for mapping, ignoring the time sequence dependence and internal geometric structure of the features, and lacking targeted modeling of the sensitivity difference of the boundaries required for different state discrimination, resulting in insufficient generalization and recognition ability for complex abnormal patterns. To this end, the application proposes a state classification method that realizes accurate and robust charging cabinet running state recognition by fusing rule-sensitive gating, spatio-temporal attention focusing, and multi-prototype metric learning. For the corresponding rule-sensitive gating, the embodiment of the application calculates the boundary rule-sensitive feature vector corresponding to the data, quantifies the abnormal risk of different channels at different times, so that the model is alert to the physical safety boundary.

[0081] In specific implementation, the out-of-bound event data corresponding to the target monitoring data can be determined based on a preset out-of-bound parameter (in an embodiment, the absolute safety boundary described above can be used as the out-of-bound parameter); and the boundary proximity data corresponding to the target monitoring data can be determined based on the normalized relative boundary corresponding to the normalization processing. Then, the boundary rule-sensitive features of the target monitoring data are determined based on the out-of-bound event data and the boundary proximity data. In combination with the above steps, the embodiment of the application uses the absolute safety boundary and the local statistics to construct the boundary rule-sensitive features, to capture explicit out-of-bound behavior and quantify the proximity of the data points to the safety boundary. Wherein, the boundary rule-sensitive features can be represented as:

[0082]

[0083] In the formula, represents the boundary rule-sensitive feature of the i-th channel at the j-th time point. ​The boundary violation sensitive feature value of the moment is a scalar, and the numerical value directly reflects the risk level of the position, and the greater the value, the higher the risk. represents the standard deviation of the th channel sequence in the local window centered at the th moment; represents the numerical stabilization increment, which is a very small positive number, used to prevent the denominator from being zero, and an example of the value is . represents the natural exponential function. represents the proximity perception weight coefficient, which is a hyperparameter used to balance the contribution of the absolute boundary violation penalty term and the boundary proximity perception term, and an example of the value is 0.5. represents the hyperbolic tangent activation function, which maps the input to the interval, and is used for smooth saturation amplification of the boundary proximity signal. represents the normalized relative boundary proximity of the th channel at the th moment, which is used to quantify the relative distance of the data point from the nearest absolute boundary, and the smaller the value, the closer to the boundary, and the calculation method is represented as:

[0084]

[0085] represents the positive distance of the th channel at the th moment to the absolute safe upper boundary, which is used to measure the degree of upper boundary violation, and the calculation method is represented as: , when , it means that the upper boundary violation occurs, which may correspond to the occurrence of the corresponding boundary event; represents the positive distance of the th channel at the th moment to the absolute safe lower boundary, which is used to measure the degree of lower boundary violation, and the calculation method is represented as: , when , it means that the lower boundary violation occurs, which may correspond to the occurrence of the corresponding boundary event. represents the value of the original data sample in the th channel and the th time point.

[0086] Step S212, the boundary violation sensitive feature is fused into the channel interaction fusion vector to construct the focused feature tensor corresponding to the target monitoring data.

[0087] The channel interaction fusion vector is a global correlation feature extracted based on the channel-time-depth cross-dimensional coupling mode, representing the collaborative change law of multiple physical quantities; the boundary violation sensitive feature is a local critical feature, representing the abnormal precursor of data approaching the safety threshold. This embodiment of the invention fuses the two, so that the corresponding features simultaneously contain feature information at both the global correlation level and the local critical level. This allows for focusing on the differential information of the two types of features, amplifying the feature differences between normal and abnormal states (especially early critical anomalies and complex collaborative anomalies), thereby achieving highly discriminative classification and prediction.

[0088] In this embodiment of the invention, a preset gating fusion mechanism is used to fuse boundary violation sensitive features into the channel interaction fusion vector to construct an initial fusion feature tensor. Based on preset multi-dimensional attention weights, the initial fusion feature tensor is adaptively weighted to construct a focused feature tensor. In one implementation, a channel- and time-adaptive gating fusion mechanism can be used to inject boundary risk perception information into the enhanced feature tensor to obtain the fusion feature tensor. This allows the model to dynamically adjust the contribution of the enhanced features based on the risk value at each location and directly inject risk information. Specifically, the calculation method of the fusion feature tensor is expressed as follows:

[0089]

[0090] In the formula, Represents the fusion feature tensor In the The first channel, the first Time, Number The feature values ​​of each deep pattern represent the deep coupling features (i.e., the initial fused feature tensor) after being modulated by boundary risk perception and injected with direct risk signals. The dimension is . This represents the above-mentioned enhanced feature tensor In the The first channel, the first Time, Number Feature values ​​of a depth pattern; This represents the element-wise multiplication operator; This represents the Sigmoid activation function, which maps the input to... The interval is used as the gating weight. Indicates the first The first channel, the first The gating transformation weights corresponding to each depth pattern are trainable parameters, which are scalars used to map boundary violation sensitive feature values ​​to gating signals. Indicates the first The gating bias corresponding to each depth pattern is a trainable parameter and is a scalar. Indicates the first the risk information injection weight corresponding to the first depth mode of the first channel, is a trainable parameter, is a scalar, and is used for linearly transforming the boundary violation sensitive feature and then injecting the enhanced feature tensor.

[0091] Further, the contribution degrees of different stages and different coupling modes of the charging cabinet operation to the final state discrimination are different, in order to focus on the most critical time segment and the most significant coupling mode, embodiments of the present application also calculate the attention weights in the time dimension and the depth dimension at the same time under the premise of not flattening the tensor, which is represented as:

[0092]

[0093] In the formula, the attention weight matrix is represented as the element of the first channel, the first time, and the first depth mode, and the numerical size represents the importance score of the corresponding feature. the attention score of the first channel, the first time, and the first depth mode is represented as:

[0094]

[0095] the query vector of the first channel and the first time is obtained by linearly transforming all the depth feature vectors of the fusion feature tensor at a specific channel and time point, and the calculation manner is represented as: . the fusion feature tensor is represented as the feature vector of the first channel, the first time, and all depth modes is represented as . ; the weight matrix of the query transformation is represented as

[0096] ​​​​​​​​​​​​​​​​

[0097] is a weight matrix representing key transformation, which is a trainable parameter, used to linearly project the feature vector aggregated after pooling into a new vector space; is the dimension of the key vector, which is set by default ; is a content enhancement coefficient, which is a trainable scalar parameter, used to adjust the degree of direct influence of the original feature value on the attention weight.

[0098] Further, the initial fusion feature tensor is reweighted using the above attention weight matrix to highlight the feature elements that contribute more to state recognition and suppress unimportant features, denoted as:

[0099]

[0100] In the formula, is a focused feature tensor is the feature value of the i-th channel, the t-th moment, and the j-th depth mode, which represents the depth-coupled feature carrying boundary risk perception after importance calibration (i.e., the focused feature tensor). Step S214, using a pre-constructed operating state monitoring model to classify and predict the focused feature tensor to determine the operating state classification result corresponding to the target charging cabinet.

[0101] After determining the corresponding focused feature tensor, the embodiment of the present application classifies and predicts based on the learned feature law through the pre-constructed model, and outputs the corresponding operating state classification result. For example, it includes normal operating state, critical early warning state, and specific abnormal type (overvoltage / overcurrent, etc.).

[0102] Considering that different operating states may correspond to multiple typical patterns in the feature space, the embodiment of the present application adopts multi-prototype metric learning for classification output. In specific implementation, the multi-prototype metric learning mechanism is adopted to determine that each monitoring channel of the focused feature tensor is respectively based on the probability distribution prediction result of the corresponding mode summary vector; based on the probability distribution prediction result, the operating state classification result corresponding to the target charging cabinet is determined. Wherein, the focused feature tensor is first aggregated into a mode summary vector of each channel, then the minimum Mahalanobis distance between the mode summary vector and multiple prototypes of each state category is calculated, and finally the prediction probability of the sample belonging to each operating state category is obtained by converting the probability distribution, denoted as:

[0103] Considering that different operating states may correspond to multiple typical patterns in the feature space, the embodiment of the present application adopts multi-prototype metric learning for classification output. In specific implementation, the multi-prototype metric learning mechanism is adopted to determine that each monitoring channel of the focused feature tensor is respectively based on the probability distribution prediction result of the corresponding mode summary vector; based on the probability distribution prediction result, the operating state classification result corresponding to the target charging cabinet is determined. Wherein, the focused feature tensor is first aggregated into a mode summary vector of each channel, then the minimum Mahalanobis distance between the mode summary vector and multiple prototypes of each state category is calculated, and finally the prediction probability of the sample belonging to each operating state category is obtained by converting the probability distribution, denoted as:

[0104]

[0105] In the formula, ​​Indicates that the sample belongs to the first The predicted probability of each running state category, with the sum of the probabilities of all categories being 1. Indicates the index of running status category. ; This indicates the total number of categories of operating status. For example, specific categories include: 1-Normal, 2-Overvoltage, 3-Undervoltage, 4-Overcurrent, 5-Power anomaly, and 6-Combined anomaly. Indicates the first The number of prototypes corresponding to each category is a hyperparameter; Indicates the prototype index. Indicates the selection of the first The operation is to find the minimum nearest distance among all prototypes in each category.

[0106] Indicates sample features and the first The first category The overall distance between prototypes is calculated as follows: ; The calculation method for the squared Mahalanobis distance is as follows: ;in It is a trainable, positive semi-definite diagonal weight matrix used to measure the difference in importance of different depth pattern dimensions. yes The transpose, and It is a column vector. Indicates the first Category 1, No. 2 The corresponding prototype in the first The prototype subvectors of each channel are trainable parameters with dimension . . Indicates the first The pattern summary vector for each channel can be obtained by averaging the focus feature tensor over the time dimension, and the calculation method is expressed as follows: , dimension . Represents the focused feature tensor In the The first channel, the first Feature vectors at all times and across all depth modes.

[0107] In summary, for the target charging cabinet of the embodiment of the present application, firstly, one or more charging cabinets are monitored in real time or offline running state, and the data collection process is started (that is, through the Internet of Things sensor network of the target charging cabinet, the current or recent current, voltage, power time sequence monitoring data is synchronously collected according to the same specification as the training stage, and a raw data sample to be detected is formed). After that, the sample is processed by the above-mentioned dynamic double boundary normalization method, and by using the pre-defined absolute safety boundary and the same local statistical calculation rule, it is converted into normalized data to pre-process the data and highlight possible boundary violation information. Further, the channel interaction fusion and boundary violation sensitive feature fusion can be fused into the above-mentioned model to classify and identify the above-mentioned normalized data through the model. The model has completed training and loaded the optimal parameters, and the model internally sequentially multi-source time sequence feature coupling tensor enhancement (that is, constructing a channel interaction fusion vector), and charging cabinet running state classification based on determination of boundary violation sensitive features.

[0108] Corresponding to the above-mentioned embodiment, the model first mines the deep coupling mode between the current, voltage and power sequence through the coupling tensor enhancement module and filters out the noise; then, through the classification module, the boundary violation sensitive features corresponding to the original data can be combined, the key information can be focused through the gating fusion and the spatiotemporal attention mechanism, and finally the probability that the to-be-detected sample belongs to each preset running state category (normal, overvoltage, undervoltage, overcurrent, power anomaly, composite anomaly) can be calculated through multi-prototype metric learning. The model output is a probability distribution vector, and the category with the highest probability is determined as the running state of the charging cabinet in the current monitoring period. The monitoring system can trigger the corresponding early warning, alarm or log record in real time according to the determination result, so as to realize the automatic and intelligent monitoring and early fault identification of the charging cabinet running state, and provide direct decision support for operation and maintenance management. Correspondingly, Figure 4 The classification prediction process schematic diagram corresponding to the embodiment of the present application is shown.

[0109] Further, in combination with the above-mentioned embodiments, the running state monitoring model of the embodiments of the present application can be constructed based on the channel interaction fusion vector and the boundary violation sensitive feature of the corresponding training sample, so that the feature representation dimension of the model changes from a single physical quantity change dimension to a two-dimensional space of global coupling + local criticality. On the one hand, the global coupling feature can ensure the model's ability to identify multiple parameter coordination anomalies and avoid the missed detection problem of single parameter monitoring; on the other hand, the local criticality feature can enhance the model's sensitivity to early fault precursors and break the after-the-fact judgment limitation of exceeding the standard to alarm. The combination of the two can make the feature space of the model have higher distinguishability to accurately distinguish normal operation, critical early warning, and specific abnormal types (overvoltage, overcurrent, etc.), and significantly improve the recognition accuracy and generalization ability of the charging cabinet running state under complex working conditions. Correspondingly, after the model training is completed, the output result can include not only the category judgment of the running state, but also the feature weight backtracking positioning of the abnormal source (such as a certain channel critical abnormality, multiple channel coordination abnormality). The model can not only realize automatic state recognition, but also provide precise abnormal positioning basis for operation and maintenance personnel.

[0110] Among them, the corresponding training sample can be constructed based on the above-mentioned monitoring data to form a standardized data set for training the charging cabinet running state monitoring model. Among them, the original time sequence monitoring sequence can be collected from a large number of normal and different abnormal state charging cabinet historical running records to construct the corresponding training data set. The collection process needs to ensure the consistency of sensor accuracy, sampling rate and strict synchronization of time stamp. After the data collection is completed, each original data sample is also accurately labeled to provide the supervision signal required for model training. The labeling work can determine and classify the running state of the charging cabinet corresponding to each time sequence data sample according to the historical fault record, operation and maintenance log and expert knowledge of the charging cabinet. In one embodiment, the defined running state categories are six categories, including: normal state, overvoltage state, under-voltage state, over-current state, power abnormal state and composite abnormal state, wherein the composite abnormal state refers to a condition that contains multiple single abnormal features or presents a complex abnormal pattern. The labeling result is given in the form of sample label, using one-hot encoding format. Finally, a large number of labeled original data samples and their corresponding one-hot encoding labels can be combined to form a complete training data set.

[0111] Furthermore, the loss function employed in this embodiment of the invention not only supervises the correctness of the final classification but also guides the model to focus more on high-risk samples and ensures the stability of feature enhancement by introducing violation-aware sample weights and constraints on the coupled-mode filtering process. In specific implementation, the loss function includes a violation-aware weighted classification loss and a coupled-mode filtering consistency regularization loss. The violation-aware weighted classification loss is calculated based on the dynamic weights of the training samples, and the dynamic weights are determined based on boundary violation-sensitive features; the coupled-mode filtering consistency regularization loss is used to constrain the consistency of the filtering parameters corresponding to the channel interaction fusion vector.

[0112] 1) Classification loss weighted by violation perception:

[0113] Cross-entropy loss is fundamental to supervised learning, but abnormal samples in charging cabinets are usually few in number and highly dangerous. To improve the model's classification accuracy for high-risk samples, a dynamic weight is assigned to each training sample based on the boundary violation sensitivity feature, thus constructing a violation-aware weighted classification loss, expressed as:

[0114]

[0115] In the formula, This represents the violation-aware weighted classification loss, which serves as the primary supervision signal during model training to ensure correct model classification. This represents the total number of samples in the training batch; Indicates the sample index; Indicates the first The dynamic weights of each sample are calculated as follows: This is used to amplify the loss contribution of high-risk samples. Indicates traversing the first... For each sample, take the maximum value of the boundary violation sensitive feature value among all channels and all time points, and use it to extract the highest violation risk value of the sample among all channels and all time points; This represents the violation weighting enhancement coefficient, a hyperparameter greater than 0, used to control the sensitivity of sample weights to the degree of violation. An example value is 1.0. Indicates the first The sample at the th The first channel, the first Boundary violation sensitive feature value at any given time; Indicates the first The true label of the sample is obtained using one-hot encoding. When the sample belongs to the... The value is 1 if the condition is met, otherwise it is 0. Indicates the first The sample belongs to the first The predicted probability of each running state category; Logarithm function, default base is natural constant.

[0116] 2) Coupling mode filter consistency regularization loss:

[0117] The adaptive singular value filter aims to denoise and focus on the main coupling mode. To avoid losing beneficial information or introducing instability due to the filter process being too drastic, the coupling mode filter consistency regularization loss is also used in the embodiment of the present application to constrain the singular values after filtering from deviating too much from the original singular values, while encouraging the sparsity of the coupling kernel tensor to improve interpretability, which is represented as:

[0118]

[0119] In the formula, The coupling mode filter consistency regularization loss is used to constrain and guide the feature enhancement process, and encourage the learned inter-channel coupling relationship to be more concise and more interpretable. The singular value consistency regularization coefficient is a hyperparameter greater than 0, used to control the penalty strength of the difference between the singular values before and after filtering, and an example of the value is 0.1; The coupling kernel sparsity regularization coefficient is a hyperparameter greater than 0, used to control the penalty strength of the absolute value of the elements of the coupling kernel tensor, and an example of the value is 0.01; The trainable coupling kernel tensor The scalar weight in the original channel , output channel and depth mode position, representing the weight of the information from the original channel transferred to the enhanced channel when generating the feature of the depth mode. By weighted summing the contributions of all original channels, information interaction and fusion between channels are realized; The channel index different from .

[0120] 3) Calculate the total loss function:

[0121] The total loss function of the model is the sum of the rule-aware weighted classification loss and the coupling mode filter consistency regularization loss, that is, the total loss function . Alternatively, the total loss function can also be calculated by weighted sum, for example, , and the weight value is a preset hyperparameter.

[0122] Further, the training dataset constructed by the foregoing steps is used as the basis for training the corresponding model, the forward propagation calculation model prediction result is calculated, and the total loss function is used as the model performance measurement standard and optimization target. The training process can use a gradient descent-based back propagation algorithm to update all trainable parameters in the model through iterative optimization. At the beginning of training, all trainable parameters in the model are first initialized. In each training iteration cycle (epoch) or each mini-batch, a batch of samples can be extracted from the training dataset, sequentially subjected to the dynamic double-boundary normalization preprocessing described above, and then input into the monitoring model for forward propagation. The forward propagation process can follow the data processing and classification output steps described above, and finally output the prediction probability of each sample belonging to each operating state category through the sequential execution of the multi-source time series feature coupling tensor enhancement and violation sensitive gate fusion, the spatio-temporal attention focusing and the multi-prototype measurement classification. Further, according to the prediction probability and the true label of the sample, the total loss function value can be calculated, which comprehensively considers the violation perception weighted classification loss and the coupled mode filtering consistency regularization loss. The gradient of the loss function with respect to all trainable parameters is calculated, and the selected optimizer (such as the Adam optimizer) is used to update these parameters according to the gradient direction and learning rate and other hyperparameters, so as to minimize the total loss function. This process is repeated until the preset stopping iteration condition is met. Further, the stopping condition can be set as the classification performance (such as accuracy) on the independent validation set no longer improves and lasts for a certain number of rounds (early stopping method), or reaches the preset maximum training iteration number. When the training is completed, the model parameters with the best performance are saved, that is, the training and parameter updating of the charging cabinet operating state monitoring model are completed.

[0123] In summary, the embodiments of the present application have the following innovations compared with the prior art:

[0124] 1. The dynamic double-boundary normalization method is used to construct a time-varying normalization boundary by combining an absolute safety boundary and local statistics, and to retain the out-of-bound values outside the reference interval, so that the model is sensitive to boundary violations and approaching behavior.

[0125] 2. The multi-source time series feature coupling tensor enhancement module is used to model the interaction between multiple physical quantities by introducing a learnable coupling kernel tensor, and to extract a robust coupled mode by combining an adaptive singular value filter, thereby solving the problem that simple feature splicing cannot mine deep spatio-temporal correlations.

[0126] 3. The classification mechanism that fuses violation perception gate, spatio-temporal double attention and multi-prototype measurement learning is used to quantify abnormal risks through boundary violation sensitive features, and to realize adaptive feature calibration through gate fusion and attention focusing, and finally to enhance the recognition ability for complex abnormal patterns through multi-prototype matching.

[0127] 4. Adopting the rule violation perception weighted loss function and the coupled mode filtering consistency regularization loss, the attention to the high-risk abnormal sample is strengthened through the sample dynamic weight, and the stability and interpretability of the feature enhancement process are constrained, and the performance of the model in the data imbalance scene is improved.

[0128] Further, on the basis of the above-mentioned embodiments, the embodiment of the present application also provides a charging cabinet operation state monitoring device based on Internet of Things data, referring to Figure 5 The device comprises: a data monitoring module 10, configured to determine the operation monitoring data of the target charging cabinet through the Internet of Things sensor network of the target charging cabinet; a data preprocessing module 20, configured to perform normalization processing on the operation monitoring data based on the preset absolute safety boundary and the local statistical features corresponding to each monitoring channel of the operation monitoring data, to generate target monitoring data; a data processing module 30, configured to construct a channel interaction fusion vector of the target monitoring data and determine a boundary violation sensitive feature of the target monitoring data; and an execution module 40, configured to perform classification prediction on the target monitoring data based on the channel interaction fusion vector and the boundary violation sensitive feature, to determine the operation state classification result of the target charging cabinet.

[0129] The charging cabinet operation state monitoring device based on Internet of Things data provided by the embodiment of the present application has the same implementation principle and technical effects as the foregoing method embodiment, and for brief description, the part not mentioned in the device embodiment can refer to the corresponding content in the foregoing method embodiment.

[0130] Further, the embodiment of the present application also provides a charging cabinet operation state monitoring system based on Internet of Things data, which is configured with the device of the above-mentioned embodiment and is used to execute the method of any one of the above-mentioned embodiments.

[0131] The charging cabinet operation state monitoring system based on Internet of Things data provided by the embodiment of the present application has the same implementation principle and technical effects as the foregoing method embodiment, and for brief description, the part not mentioned in the system embodiment can refer to the corresponding content in the foregoing method embodiment.

[0132] The embodiment of the present application also provides an electronic device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method of any one of the above-mentioned embodiments. Figures 1-2 The embodiment of the present application also provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by the processor to execute the steps of the method of any one of the above-mentioned embodiments. Figures 1-2 The embodiment of the present application also provides a structural schematic diagram of an electronic device, as shown in Figure 6As shown, a structural schematic diagram of the electronic device is shown, wherein the electronic device comprises a processor 101 and a memory 100, the memory 100 stores computer executable instructions capable of being executed by the processor 101, and the processor 101 executes the computer executable instructions to implement the above Figures 1-2 any of the methods shown.

[0133] In Figure 6 In the embodiment shown, the electronic device further comprises a bus 102 and a communication interface 103, wherein the processor 101, the communication interface 103 and the memory 100 are connected through the bus 102. The memory 100 can contain a high-speed random access memory (RAM), and can also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 103 (which can be wired or wireless), and the Internet, a wide area network, a local network, a metropolitan area network, etc. can be used. The bus 102 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc., and can also be an AMBA (Advanced Microcontroller Bus Architecture) bus, wherein AMBA defines three kinds of buses, including an APB (Advanced Peripheral Bus) bus, an AHB (Advanced High-performance Bus) bus and an AXI (Advanced eXtensible Interface) bus. The bus 102 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6only one bus or one type of bus. The processor 101 can be a chip that, among other things, processes information. The processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), or the like. In the various embodiments, the processor 101 can be a combination of general-purpose processors, including a central processing unit (CPU), a network processor (NP), or the like. The general-purpose processor can be a microprocessor or the processor can be any conventional processor, controller, microcontroller, or state machine. Software modules (e.g., 106) can be stored on the storage media 105, memory or disk (e.g., 107), or any other appropriate storage medium or device(s) operated on or in communication with the processor 101. The processor 101 executes instructions, codes, or Figures 1-2 any of the methods disclosed.

[0134] The computer program product of the charging cabinet operation state monitoring method, device and system based on Internet of Things data provided by the embodiment of the application comprises a computer readable storage medium storing program codes, the instructions included in the program codes can be used to execute the method described in the foregoing method embodiment, and the specific implementation can be referred to the method embodiment, which will not be described here. The skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the foregoing method embodiment, which will not be described here. In addition, in the description of the embodiment of the application, unless otherwise explicitly specified and limited, the terms “installation”, “connection”, “connection” should be understood in a broad sense, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, or the communication between two elements. For the skilled in the art, the specific meaning of the above terms in the application can be understood according to the specific circumstances. If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the application or the part of the technical solutions that essentially contribute to the prior art can be embodied in the form of software products, and the computer software product is stored in a storage medium, including a plurality of instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the application. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and various media that can store program codes.

[0135] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance. Finally, it should be noted that: the above examples are only specific embodiments of the present application, used to illustrate the technical solutions of the present application, and are not a limitation on it, the protection scope of the present application is not limited to this, although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand: any person skilled in the art within the technical range disclosed by the present application, it can still modify or easily think of changes to the technical solutions recorded in the foregoing examples, or make equivalent replacement to part of the technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for monitoring the operating state of an Internet of Things data-based charging cabinet, characterized in that, The method comprises: determining operation monitoring data of the target charging cabinet through an Internet of Things sensor network of the target charging cabinet; based on a preset absolute safety boundary and a local statistical feature corresponding to each monitoring channel of the operation monitoring data, respectively, normalizing the operation monitoring data to generate target monitoring data; constructing a channel interaction fusion vector of the target monitoring data and determining a boundary violation sensitive feature of the target monitoring data; based on the channel interaction fusion vector and the boundary violation sensitive feature, classifying and predicting the target monitoring data to determine the operation state classification result of the target charging cabinet; wherein, based on the channel interaction fusion vector and the boundary violation sensitive feature, classifying and predicting the target monitoring data to determine the operation state classification result of the target charging cabinet, the step comprises: fusing the boundary violation sensitive feature into the channel interaction fusion vector to construct a focused feature tensor corresponding to the target monitoring data; using a pre-constructed operation state monitoring model to classify and predict the focused feature tensor to determine the operation state classification result corresponding to the target charging cabinet; the step of constructing the channel interaction fusion vector of the target monitoring data comprises: using a preset coupling kernel tensor to fuse the inter-channel information of the monitoring sequence of each monitoring channel of the target monitoring data to generate an intermediate tensor corresponding to the target monitoring data; performing singular value decomposition on the intermediate tensor to determine a first change mode for the time joint dimension and a second change mode for the data depth dimension of the target monitoring data; based on a first singular vector matrix corresponding to the first change mode and a second singular vector matrix corresponding to the second change mode, performing data filtering processing on the target monitoring data to generate the channel interaction fusion vector of the target monitoring data; the step of determining the boundary violation sensitive feature of the target monitoring data comprises: based on a preset out-of-bound parameter, determining out-of-bound event data corresponding to the target monitoring data; based on a normalized relative boundary corresponding to the normalization processing, determining boundary proximity data corresponding to the target monitoring data; based on the out-of-bound event data and the boundary proximity data, determining the boundary violation sensitive feature of the target monitoring data.

2. The method of claim 1, wherein, based on a preset absolute safety boundary and a local statistical feature corresponding to each monitoring channel of the operation monitoring data, respectively, normalizing the operation monitoring data to generate target monitoring data, the step comprises: based on a preset local window length, determining a local average level corresponding to each monitoring channel of the operation monitoring data; based on a preset absolute safety boundary, determining a dynamic normalization boundary corresponding to the local average level; based on the dynamic normalization boundary, normalizing the monitoring sequence of each monitoring channel, respectively, to construct the target monitoring data corresponding to the operation monitoring data.

3. The method of claim 1, wherein, The step of fusing the boundary violation sensitive feature into the channel interaction fusion vector to construct the focused feature tensor corresponding to the target monitoring data comprises: The preset gating fusion mechanism is adopted to fuse the boundary violation sensitive feature into the channel interaction fusion vector to construct an initial fusion feature tensor; The initial fusion feature tensor is adaptively weighted based on the preset multi-dimensional attention weight to construct a focused feature tensor.

4. The method of claim 1, wherein, The step of using the pre-constructed operation state monitoring model to perform classification prediction on the focused feature tensor to determine the operation state classification result corresponding to the target charging cabinet comprises: A multi-prototype metric learning mechanism is adopted to determine that each monitoring channel of the focused feature tensor is based on the probability distribution prediction result of the corresponding mode summary vector; Based on the probability distribution prediction result, the operation state classification result corresponding to the target charging cabinet is determined.

5. The method of claim 1, wherein, The operation state monitoring model is constructed based on the channel interaction fusion vector and the boundary violation sensitive feature of the preset training sample; The loss function of the operation state monitoring model comprises a violation perception weighted classification loss and a coupled mode filtering consistency regularization loss; The violation perception weighted classification loss is calculated based on the dynamic weight of the training sample, and the dynamic weight is determined based on the boundary violation sensitive feature; The coupled mode filtering consistency regularization loss is used to constrain the consistency of the filtering parameters corresponding to the channel interaction fusion vector.

6. A charging cabinet operation state monitoring device based on Internet of Things data, characterized in that, The device comprises: A data monitoring module configured to determine operation monitoring data of a target charging cabinet through an Internet of Things sensor network of the target charging cabinet; A data preprocessing module configured to perform normalization processing on the operation monitoring data based on a preset absolute safety boundary and local statistical features corresponding to a monitoring sequence of each monitoring channel of the operation monitoring data to generate target monitoring data; A data processing module configured to construct a channel interaction fusion vector of the target monitoring data and determine a boundary violation sensitive feature of the target monitoring data; An execution module configured to perform classification prediction on the target monitoring data based on the channel interaction fusion vector and the boundary violation sensitive feature to determine an operation state classification result of the target charging cabinet; The execution module is further configured to fuse the boundary violation sensitive feature into the channel interaction fusion vector to construct a focused feature tensor corresponding to the target monitoring data, and use a pre-constructed operation state monitoring model to perform classification prediction on the focused feature tensor to determine the operation state classification result corresponding to the target charging cabinet. The data processing module is further configured to: perform inter-channel information fusion on a monitoring sequence of each monitoring channel of the target monitoring data by using a preset coupling nuclear tensor, to generate an intermediate tensor corresponding to the target monitoring data; perform singular value decomposition on the intermediate tensor to determine a first variation mode of the target monitoring data with respect to a time joint dimension and a second variation mode of the target monitoring data with respect to a data depth dimension; and perform data filtering processing on the target monitoring data based on a first singular vector matrix corresponding to the first variation mode and a second singular vector matrix corresponding to the second variation mode, to generate a channel interaction fusion vector of the target monitoring data. The data processing module is further configured to: determine out-of-bound event data corresponding to the target monitoring data based on a preset out-of-bound parameter; determine boundary proximity data corresponding to the target monitoring data based on a normalized relative boundary corresponding to the normalization processing; and determine a boundary violation sensitive feature of the target monitoring data based on the out-of-bound event data and the boundary proximity data.

7. An Internet of Things data-based charging cabinet operation state monitoring system, characterized in that, The system is configured with the device of claim 6 to perform the method of any one of claims 1-5.

Images