Electric control cabinet fault early warning method and system based on deep learning

Abstract

The invention relates to the technical field of electrical equipment state monitoring and fault diagnosis, and discloses an electrical control cabinet fault early warning method and system based on deep learning, and the method comprises the steps: obtaining physical sensing data and control logic data, and carrying out the time sequence alignment; parsing the control data to identify a working condition mode; in the self-calibration mode, the actual temperature drop rate is used for updating a correction coefficient representing heat dissipation attenuation of the equipment; in a monitoring mode, effective thermodynamic parameters are reconstructed through a depth parameter estimation network and a correction coefficient; reversely resolving the real-time contact impedance of the monitoring node based on a heat balance differential equation; and monitoring an impedance evolution trend, and generating an early warning signal when the impedance monotonically rises and is decoupled from a load current change trend. According to the invention, through fusing the physical mechanism model and the data driving network, accurate monitoring and fault diagnosis of the contact impedance under a complex thermal environment and an equipment aging condition are realized.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment condition monitoring and fault diagnosis technology, specifically to a method and system for early warning of electrical control cabinet faults based on deep learning. Background Technology

[0002] As the power distribution and control hub of a power system, the electrical control cabinet contains a dense array of circuit breakers, contactors, and various terminal blocks. During long-term operation, factors such as mechanical vibration, thermal expansion and contraction, and surface oxidation can cause electrical connections to loosen or deteriorate, leading to abnormally high contact resistance. If such faults are not detected in time, the continuous accumulation of Joule heat can cause insulation aging, equipment burnout, and even electrical fires.

[0003] Currently, monitoring of electrical contact faults mainly relies on temperature detection methods, such as wireless temperature sensors or infrared thermal imaging technology. However, node temperature is the result of the combined effects of load current, contact resistance, ambient temperature, and heat dissipation conditions, and relying solely on a single temperature threshold has limitations. Under heavy load conditions, even normal electrical connections may generate high temperature rises, easily triggering false alarms; while under light load or forced air cooling conditions, even with poor contact, the node temperature may not reach the preset threshold, leading to missed alarms.

[0004] To obtain a more fundamental understanding of the contact state, some existing technologies attempt to invert the contact resistance using the principle of thermal equilibrium. This method requires establishing an accurate thermodynamic model, where the convective heat transfer coefficient and the external heat flux are key boundary parameters that determine the accuracy of the calculation.

[0005] However, in actual industrial scenarios, the thermal environment of electrical control cabinets is extremely complex and dynamically changing.

[0006] On the one hand, the starting and stopping of the cooling fan, the accumulation of dust on the dust filter, and the aging and deterioration of the fan itself can all cause the convective heat transfer capacity to change over time. On the other hand, the heat generated by the operation of adjacent equipment inside the cabinet will produce radiative thermal interference that is difficult to measure. Existing modeling methods usually simplify these thermodynamic parameters to fixed constants or simple empirical formulas, which cannot track changes in heat dissipation conditions and interference from external thermal fields in real time. This results in the calculated contact impedance value containing a large number of error components caused by environmental factors, making it difficult to accurately reflect the true physical state of the contact surface and limiting the practical application effect of this technology under complex working conditions. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for early warning of electrical control cabinet faults based on deep learning, which solves the problems of low calculation accuracy and high false alarm rate caused by the susceptibility of existing contact impedance monitoring technology to fluctuations in ambient temperature, changes in heat dissipation conditions and dynamic load interference.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based fault early warning method for electrical control cabinets, comprising the following steps: First, the system performs data acquisition and timing alignment, acquiring physical sensor data and control logic data from within the electrical control cabinet.

[0009] Among them, physical sensor data reflects the real-time operating status of electrical circuits, while control logic data reflects the system's active heat dissipation behavior and the status of external thermal interference sources. By aligning these two types of heterogeneous data through a unified clock source, a spatiotemporally consistent data foundation is provided for subsequent multi-source information fusion.

[0010] Secondly, the system identifies the current operating mode based on control logic data. This operating mode is divided into self-calibration mode and monitoring mode. The system makes logical judgments based on the fan operating status and circuit current level, realizing adaptive perception of different operating stages of the equipment.

[0011] In self-calibration mode, the system performs online correction of thermodynamic parameters. It uses the temperature change characteristics of the monitoring node when there is no current load and the cooling fan is running to calculate the actual temperature drop rate of the monitoring node and compares it with the theoretical temperature drop rate calculated based on the design parameters and Newton's law of cooling.

[0012] The heat dissipation correction coefficient is updated by comparing the results. This coefficient quantitatively characterizes the degree of attenuation of the actual heat dissipation capacity of the electrical control cabinet relative to the initial design value due to filter blockage, fan aging, or dust accumulation in the air duct. This step ensures that the thermal model can adaptively adjust as the equipment ages, eliminating calculation errors caused by model parameter drift.

[0013] In monitoring mode, the system performs parameter reconstruction based on deep learning, mapping control logic data including fan commands and the load status of nearby devices into feature vectors, which are then input into a preset deep parameter estimation network.

[0014] The network outputs nominal thermodynamic parameters (including nominal convective heat transfer coefficient and external disturbance heat flux) through nonlinear mapping, and combines the aforementioned updated heat dissipation correction coefficient to perform weighted correction on the nominal parameters, thereby obtaining effective thermodynamic parameters that conform to the current actual operating conditions. This process realizes real-time and accurate estimation of thermodynamic boundary conditions that are difficult to measure directly under complex and variable operating conditions.

[0015] Subsequently, the system performs a physical reverse decoupling calculation of the contact impedance, and constructs a differential equation based on the principle of unsteady thermal equilibrium, which includes contact impedance terms, thermal inertia terms, convection heat dissipation terms, and external disturbance terms.

[0016] By substituting the real-time acquired current value, node temperature value, and reconstructed effective thermodynamic parameters into the inverse form of the differential equation, and through numerical calculation, the temperature rise component caused by ambient temperature, convective heat dissipation, and external heat source is extracted from the total temperature rise data, and the real-time contact impedance value caused solely by the Joule heating effect of contact resistance is calculated.

[0017] Finally, the system performs impedance evolution trend analysis and fault early warning. It performs statistical analysis on the calculated real-time contact impedance sequence and monitors its time evolution trend. The system introduces a current-impedance decoupling judgment mechanism. Only when the real-time contact impedance shows an irreversible monotonically increasing trend and this increasing trend is not related to the fluctuation characteristics of the load current (i.e., decoupling), it determines that there is physical contact degradation at the monitoring node and generates graded fault early warning signals.

[0018] Furthermore, in one specific implementation, the process of updating the heat dissipation correction coefficient utilizes the natural cooling or forced air cooling process during equipment shutdown or low load periods. The system collects the temperature decay curve during this stage and uses physical laws to back-calculate the current actual comprehensive heat dissipation efficiency, thereby achieving self-healing of model parameters without increasing additional hardware costs.

[0019] Furthermore, in one specific implementation, the input features of the deep parameter estimation network cover the fan start / stop status, ambient temperature, and thermal potential distribution of nearby devices. By learning the implicit correlation between control commands and thermal field distribution, the network can predict the radiative heat flux generated by the heating of nearby devices, thereby subtracting this part of the external thermal interference from the heat balance equation and improving the purity of impedance calculation.

[0020] Furthermore, in one specific implementation, the calculation process of the real-time contact impedance introduces numerical stability processing. By introducing regularization parameters and non-negative physical constraints into the denominator of the calculation formula, the numerical singularity and non-physical negative value problems caused by the zero-crossing point of the load current or sensor noise are solved, ensuring the robustness of the algorithm across the entire operating range.

[0021] A second aspect of the present invention provides a deep learning-based fault early warning system for electrical control cabinets, the system comprising: The data synchronization module is configured to acquire physical sensor data and control logic data, and perform data cleaning and timing alignment processing. The state mapping module is configured to parse control logic data and identify whether the system is currently in self-calibration mode or monitoring mode. The parameter calibration module is configured to update the heat dissipation correction coefficient based on the actual temperature drop data of the monitoring node in self-calibration mode, so as to realize the quantitative assessment of the aging status of the heat dissipation system. The parameter reconstruction module is configured to run a deep parameter estimation network in monitoring mode to obtain nominal thermodynamic parameters and generate effective thermodynamic parameters using a heat dissipation correction factor. Impedance calculation module, configured to build an unsteady thermal equilibrium model, uses physical sensor data and effective thermodynamic parameters to inversely calculate the real-time contact impedance of the monitoring node; The trend analysis module is configured to monitor the evolution characteristics of impedance values ​​and perform fault determination and early warning generation based on the monotonicity of impedance trends and the decoupling characteristics from current fluctuations.

[0022] This invention provides a method and system for early warning of electrical control cabinet faults based on deep learning. It has the following beneficial effects: 1. This invention achieves accurate calculation of contact impedance under complex thermal environments by combining a deep parameter estimation network with thermal balance differential equations. This method utilizes a deep network to process the nonlinear relationship between control commands and thermal field distribution, reconstructing external heat flux and convective heat transfer coefficients that are difficult to measure directly. By utilizing the law of conservation of energy in physical equations, the influence of ambient temperature fluctuations and external heat source radiation is eliminated from the total temperature rise data. Compared with traditional monitoring methods that rely solely on surface temperature thresholds, this invention can obtain impedance values ​​that reflect the true state of the contact surface under dynamic load changes and environmental interference without adding invasive sensors.

[0023] 2. This invention establishes a parameter adaptive calibration mechanism based on operating mode switching, which solves the problem of thermal model mismatch caused by dust accumulation or component aging during long-term operation. The system uses temperature drop data during low load or downtime of the equipment to infer the current actual heat dissipation capacity and update the correction coefficient. It can automatically compensate for thermal parameter drift caused by dust filter blockage, fan performance degradation or increased air duct resistance, ensuring that the fault early warning model maintains calculation stability throughout the complete service life of the electrical control cabinet, without the need for frequent manual parameter calibration.

[0024] 3. This invention introduces a decoupling judgment mechanism between impedance evolution trend and load current, which significantly improves the anti-interference capability of fault early warning. By statistically analyzing the monotonicity of impedance change and its correlation with current fluctuation, the system can effectively identify and shield false signals caused by current transformer measurement errors, thermal model transient response lag, or load mutations. This logic ensures that the system only triggers alarms when substantial physical connection degradation occurs at the monitoring node, reducing the false alarm rate and providing maintenance personnel with a highly reliable diagnostic basis. Attached Figure Description

[0025] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention.

[0026] Among them, 10 is the monitoring node; 20 is the current transformer; 30 is the temperature sensor; 31 is the environmental sensor; 40 is the controller; 50 is the computing gateway; 51 is the data synchronization module; 52 is the state mapping module; 53 is the parameter reconstruction module; 54 is the impedance calculation module; 55 is the parameter calibration module; 56 is the trend analysis module; and 100 is the electrical control cabinet. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example: Please see the appendix Figure 1 This invention provides a deep learning-based fault early warning system for electrical control cabinets. The system is applied to an electrical control cabinet 100, which has at least one monitoring node 10 deployed within it. The monitoring node 10 is a connecting component in an electrical circuit.

[0029] It includes a current transformer 20, which is installed on the electrical circuit where the monitoring node 10 is located, and is used to collect load current data.

[0030] The system includes a temperature sensor 30, which is disposed on or near the surface of the monitoring node 10 to collect node temperature data. The system also includes an environmental sensor 31, which is disposed inside the electrical control cabinet 100 to collect ambient temperature data inside the cabinet.

[0031] The system includes a controller 40, which is connected to the actuators within the electrical control cabinet 100. The actuators include a cooling fan and adjacent electrical equipment. The controller 40 generates a sequence of control commands, which includes the operating status of the cooling fan and the load status of the adjacent electrical equipment.

[0032] It includes a computing gateway 50, which establishes communication connections with a current transformer 20, a temperature sensor 30, an environmental sensor 31, and a controller 40. The computing gateway 50 is internally configured with a data synchronization module 51, a state mapping module 52, a parameter reconstruction module 53, an impedance calculation module 54, a parameter calibration module 55, and a trend analysis module 56.

[0033] The data synchronization module 51 is used to receive load current data, node temperature data, ambient temperature data and control command sequence, and to perform timestamp alignment processing on the data.

[0034] The state mapping module 52 is used to parse the control command sequence, convert the command signal into a control state vector, and identify the system operating mode according to the control command sequence. The operating mode includes self-calibration mode and monitoring mode.

[0035] The parameter calibration module 55 is used to operate in self-calibration mode. The parameter calibration module 55 calculates the ratio of the actual temperature drop rate to the theoretical temperature drop rate of the monitoring node 10 and generates a heat dissipation correction coefficient. The heat dissipation correction coefficient characterizes the degree of attenuation of the actual heat dissipation capacity of the electrical control cabinet 100 relative to the design value.

[0036] The parameter reconstruction module 53 is used to operate in monitoring mode. Based on the control state vector and ambient temperature data, the parameter reconstruction module 53 uses a preset parameter estimation network to output nominal thermodynamic parameters and calculates effective thermodynamic parameters in combination with the heat dissipation correction coefficient.

[0037] Impedance calculation module 54 is used to construct the thermal balance equation of monitoring node 10. Impedance calculation module 54 substitutes the load current data, node temperature data and effective thermodynamic parameters into the inverse form of the thermal balance equation to calculate the real-time contact impedance of monitoring node 10.

[0038] The trend analysis module 56 is used to monitor the time evolution characteristics of real-time contact impedance and generate an early warning signal when the real-time contact impedance shows an irreversible monotonically increasing trend.

[0039] Please see the appendix Figure 2 This invention provides a deep learning-based method for early warning of electrical control cabinet faults, comprising the following steps: S10, Data synchronization module 51 acquires physical sensor data and control logic data. The physical sensor data includes current value, node temperature value and ambient temperature value, and the control logic data includes fan control command and nearby device load command. Data synchronization module 51 aligns the physical sensor data and control logic data according to a unified clock source.

[0040] S20, the state mapping module 52 parses the control logic data to identify the current operating mode. If the control logic data indicates that the system is in a low current and forced heat dissipation is enabled, it is determined to be in self-calibration mode. If the control logic data indicates that the system is in a loaded operating state, it is determined to be in monitoring mode.

[0041] S30. When in self-calibration mode, parameter calibration module 55 calculates the temperature change rate of monitoring node 10 at the current moment. Parameter calibration module 55 compares the temperature change rate with the theoretical rate calculated based on Newton's law of cooling and updates the heat dissipation correction coefficient.

[0042] S40. When in monitoring mode, the parameter reconstruction module 53 maps the control logic data into feature vectors and inputs the feature vectors into the parameter estimation network to obtain the nominal convective heat transfer coefficient and external heat flux.

[0043] The parameter reconstruction module 53 uses the latest heat dissipation correction coefficient to perform weighted correction on the nominal convective heat transfer coefficient to obtain the effective convective heat transfer coefficient.

[0044] S50 and impedance calculation module 54 establish a differential equation containing contact impedance terms based on the principle of energy conservation. The impedance calculation module 54 substitutes the current value, node temperature value, effective convective heat transfer coefficient and external heat flux into the differential equation to calculate the real-time contact impedance of monitoring node 10.

[0045] S60, the trend analysis module 56 calculates the drift of the real-time contact impedance relative to the reference impedance. If the impedance drift continues to increase and is decoupled from the trend of current value change, the trend analysis module 56 outputs a fault warning signal.

[0046] In steps S10 and S20, the focus is on performing time-series synchronization processing and state-space mapping of multi-source data.

[0047] Step S10 specifically includes the following sub-steps: S11, Data Acquisition and Buffering. The computing gateway reads the digital sequence converted from analog signals from the current transformer and temperature sensor at a fixed high-frequency sampling rate through the first communication interface. The sampling frequency is set to effectively capture transient changes in load current and inflection points of temperature rise, for example, 1kHz to 10kHz. This data is represented as a continuous, numerically dense stream of physical sensor data. Simultaneously, the computing gateway listens to the status register address of the programmable logic controller through the second communication interface.

[0048] Since control logic data is generated only when the instruction state changes, this part of the data is represented as a time-discrete, non-equal-interval event-driven data stream.

[0049] S12, timestamp alignment of heterogeneous data. The computing gateway establishes a unified system clock reference based on the sampling time of the physical sensor data. Using the main axis, interpolation and alignment are performed on discrete control logic data.

[0050] Because control commands have state-preserving characteristics, for any sampling time If no new control event occurs at that moment, the state value corresponding to the most recent historical control event on the timeline will be used.

[0051] This zero-order hold process fills discrete event data into a continuous state sequence that strictly corresponds to the time axis of the physical sensing data.

[0052] S13. Construct the state-space vector. Based on the synchronized data, construct the physical observation vector. With control state vector .

[0053] Physical observation vector Defined as: ; in, express The effective value of the load current flowing through the monitoring node at all times; express Continuously monitor the measured surface temperature of the nodes; express The ambient background temperature inside the electrical control cabinet at all times.

[0054] This represents the first derivative of the temperature at the monitoring node with respect to time, i.e., the rate of temperature change.

[0055] This temperature change rate characterizes the speed at which nodal thermal energy accumulates or dissipates, specifically through the measurement of... The historical sequence is calculated using the five-point central difference algorithm or the sliding window linear regression algorithm to eliminate numerical jitter caused by single-point measurement noise.

[0056] Control State Vector Defined as: ; in, This is the operating status variable for the cooling fan. It takes a value of 1 when the fan is running at full speed and a value of 0 when it is stopped. If it is a speed-controlled fan, it takes its duty cycle command value.

[0057] Indicates the first A load rate command for a neighboring device that causes thermal interference to the monitoring node.

[0058] It should be noted that which devices are considered adjacent devices are determined in advance based on the physical layout of the electrical control cabinet. During the system initialization phase, a thermal topology mapping table is pre-set. This table records the high-power heat-generating devices within a set radius around the monitoring node and their corresponding PLC register addresses. The calculation gateway extracts the corresponding load rate instructions based on this mapping table and fills them into the vector.

[0059] Step S20 specifically includes the following sub-steps: S21, Operating condition feature extraction. The state mapping module reads the current physical observation vector. Current value in and control state vector Fan status .

[0060] S22, Self-calibration mode determination. This step aims to identify a physical window free from internal Joule heat sources and with well-defined heat dissipation conditions for calibrating the system's thermal damping characteristics.

[0061] The system sets the current dead zone threshold. This threshold is determined based on the zero-point drift range of the current transformer or a very small percentage of the rated current of the monitoring node.

[0062] When the system detects that the load current value is less than the current dead zone threshold, and at the same time detects that the fan status variable indicates that the forced cooling system is in the on state, the current operating condition is determined to be self-calibration mode. In this mode, the temperature change of the monitoring node is completely determined by the temperature difference between it and the environment and the heat dissipation conditions, eliminating the interference of contact resistance heating.

[0063] S23, Monitoring Mode Determination. When the system detects that the load current value is greater than or equal to the current dead zone threshold, it determines that the current operating condition is in monitoring mode. In this mode, the monitoring node exhibits a significant Joule heating effect caused by the current, satisfying the physical premise of contact impedance inverse solution.

[0064] S24, Mode Switching Hold. To avoid frequent mode switching when current fluctuations pass through the threshold critical point, a time lag is introduced into the discrimination logic. The state mapping module only executes the mode state machine switching operation after the operating conditions have been continuously met for a preset stable duration (e.g., maintained for more than 5 seconds).

[0065] In step S30, an adaptive calibration of thermal damping based on a non-working window is performed. This step utilizes a specific operating window of the electrical equipment under low load and with forced cooling to correct the boundary condition parameters of the thermophysical model online.

[0066] Step S30 specifically includes the following sub-steps: S31, Theoretical heat dissipation baseline calculation. When the system is in self-calibration mode, the parameter calibration module retrieves the physical property parameters of the monitoring nodes, including the equivalent heat capacity of the nodes. Effective heat dissipation area And the design convective heat transfer coefficient corresponding to the current fan command. .

[0067] Among them, equivalent heat capacity It is determined in advance by multiplying the specific heat capacity of the metal conductor of the monitoring node (such as the specific heat capacity of copper or aluminum) with its mass; Design convective heat transfer coefficient It is a lookup table built based on fluid simulation data or laboratory calibration data of electrical cabinet duct design, and retrieved according to the current fan speed command.

[0068] Under this operating condition, since the Joule heat generated by the load current is close to zero and the external disturbance heat flux is at a low level, the thermal balance equation of the monitoring node degenerates into a pure heat dissipation process. The parameter calibration module, based on Newton's law of cooling, calculates the theoretical temperature drop rate that the monitoring node should exhibit under the current temperature difference, assuming the heat dissipation system meets design standards. : ; in, To monitor the excess temperature of a node relative to its environment, the negative sign in the formula indicates that when there is no internal heat source and the node temperature is higher than the ambient temperature, the temperature decreases monotonically over time.

[0069] S32, Instantaneous heat dissipation performance comparison. The parameter calibration module reads the actual temperature change rate of the monitoring node calculated in step S10. The parameter calibration module compares the actual temperature change rate with the theoretical temperature drop rate to calculate the instantaneous heat dissipation efficiency coefficient. : ; The calculation process is set with a temperature difference trigger threshold. (For example, set to 5 degrees Celsius), only when the actual excess temperature At this time, the system performs a division operation. This threshold is set based on the noise floor of the temperature sensor, designed to prevent minute temperature fluctuations from causing the denominator to approach zero when the temperature difference is too small, thus leading to divergent or severely distorted calculation results. If the current excess temperature is below this threshold, the system pauses updates and retains the value from the previous moment.

[0070] S33, Smooth update of the correction coefficient. The parameter calibration module uses an exponentially weighted moving average algorithm to update the heat dissipation performance correction coefficient. Iterative updates are performed to suppress random errors in a single measurement. The update logic is as follows: ; in, The smoothing factor, with a value range of 0.95 to 0.99, determines the system's weight in remembering historical states. The closer the value is to 1, the stronger the system's anti-interference ability, but the slower its response speed to sudden environmental changes. These are the correction coefficients stored from the previous time step. Updated... It characterizes the rate of decrease in actual heat dissipation capacity relative to design capacity under the current physical environment.

[0071] S34. Parameter Holding and Anomaly Monitoring. When the system detects that the self-calibration mode conditions are no longer met, the parameter calibration module stops calculation and... Locked. In addition, the system has a built-in maintenance early warning threshold. (For example, 0.6). If calculated multiple times consecutively... If the value is below this threshold, it indicates that the heat dissipation efficiency has deviated significantly from the design value. The system generates a maintenance prompt signal for the heat dissipation system, indicating that the filter may be clogged or the fan performance may have deteriorated.

[0072] This mechanism ensures that the physical parameters used for impedance calculation in subsequent steps are adaptively corrected for operating conditions, thereby decoupling environmental factors from electrical fault factors. In this way, the system can simultaneously achieve contact fault early warning and monitor the health of the heat dissipation system.

[0073] In step S40, a dynamic thermophysical field reconstruction based on control commands is performed. In monitoring mode, this step uses the system boundary conditions determined by the control commands to reconstruct the theoretical thermodynamic parameters under the current operating conditions, providing the necessary coefficient inputs for subsequent solving of the physical equations.

[0074] Step S40 specifically includes the following sub-steps: S41, Construct the thermal interference topology mapping. The parameter reconstruction module reads the control state vector generated in step S10. For the neighboring device load rate command contained in the vector, the parameter reconstruction module introduces a pre-set thermally coupled weight vector. The elements in this weight vector Corresponding to the The radiative thermal influence factor of each neighboring device on the monitoring node is determined based on the Stefan-Boltzmann law and the geometric perspective factor. The specific calculation principle is as follows: The parameter reconstruction module performs a weighted summation of the load rate commands of the neighboring devices, which is directly proportional to the rated power of the neighboring devices and inversely proportional to the square of the geometric distance from the neighboring devices to the monitoring node, to generate a comprehensive external thermal potential characteristic. This process reduces the multidimensional discrete device control commands to a single physical characteristic quantity that characterizes the intensity of the local thermal environment of the monitoring node.

[0075] S42, Perform deep parameter estimation. The parameter reconstruction module inputs a feature vector containing fan state, weighted comprehensive external thermal potential features, and ambient temperature into the deep parameter estimation network. The deep parameter estimation network is a pre-trained multi-layer neural network model.

[0076] The physical principle behind introducing this network is as follows: The forced convection heat transfer coefficient inside the electrical cabinet has a complex nonlinear fluid dynamic relationship with fan speed, air density and duct structure, which is difficult to express with simple analytical formulas.

[0077] The network output consists of the nominal thermodynamic parameters that the monitoring nodes should possess under the current control state, specifically including the nominal convective heat transfer coefficient. and external disturbance heat flux .

[0078] This mapping process is represented as follows: ; in, Represents the forward propagation function of a neural network. This represents the network weight parameters. It's important to note that during the offline training phase, the network's input data consists of historical control command sequences, and its output labels originate from computational fluid dynamics simulation data based on a 3D model of the electrical cabinet or standard thermal test data under ideal clean conditions. Therefore, this network learns the physical mapping laws under ideal design conditions.

[0079] S43, adaptive correction of parameters under operating conditions. The parameter reconstruction module retrieves and maintains the heat dissipation performance correction coefficient updated in step S30. This coefficient, as a scalar factor characterizing the degradation of the physical environment, is used to linearly correct the nominal parameters output by the deep parameter estimation network.

[0080] The parameter reconstruction module calculates the effective convective heat transfer coefficient that is ultimately substituted into the physical equations. : ; Through this correction step, the system achieves alignment between the ideal model and actual operating conditions. For example, when the control command instructs the fan to run at full speed ( (outputs high values), but Indicator filter clog ( When corrected It will automatically decrease.

[0081] This ensures that the heat dissipation parameters used in subsequent steps accurately reflect the current physical heat dissipation capacity, preventing the temperature rise from being incorrectly attributed to contact resistance due to an overestimation of the heat dissipation capacity. This depends specifically on whether the external heat source transfers heat primarily through radiation (uncorrected) or convection (corrected).

[0082] In step S50, a physical reverse decoupling calculation of the contact impedance is performed. This step is based on the unsteady thermal balance principle of the lumped parameter method, constructs a white-box physical model of the monitoring node, and uses a numerical reverse solution algorithm to extract the Joule heat component generated by the contact resistance from the real-time temperature rise data.

[0083] Step S50 specifically includes the following sub-steps: S51, construct the unsteady-state thermal equilibrium equation. The impedance calculation module, based on the first law of thermodynamics, establishes a differential form of the energy conservation equation for the monitoring node. The physical meaning of this equation is: The increase in internal energy of a monitoring node per unit time is strictly equal to the algebraic sum of input energy and output energy, where input energy includes Joule heat generated by current flowing through contact resistance and interference heat radiated or conducted to the node from the external environment. The output energy is the heat dissipated by the node to the surrounding air through surface convection.

[0084] The specific differential equation expression is as follows: ; in, The equivalent heat capacity of the monitoring node; To monitor the actual temperature change rate of the nodes; This is the effective value of the load current; The real-time contact impedance to be solved; External disturbance heat flux; For effective convective heat transfer coefficient; The effective heat dissipation surface area of ​​the monitoring node; and These are the node temperature and the ambient temperature, respectively.

[0085] S52, Inverse analysis of contact impedance. The impedance solution module transforms the above differential equation into a equation about... The algebraic equation, through rearranging terms, will contain the parameters to be solved. The Joule heating term is separated and used as the left-hand side of the equation; the thermal inertia term ( Convection heat dissipation items ( ) and external thermal interference items ( Move it to the right side of the equation as a known excitation term.

[0086] S53, Numerical Stability Handling and Solution. Considering that the current may experience zero-crossing points or slight fluctuations in actual operation, direct division would lead to numerical instability. The impedance solution module introduces a regularization parameter. and minimum calculated current threshold .

[0087] The revised formula for calculating real-time contact impedance is as follows: ; in, It is a very small positive real number (e.g. This is used to prevent the denominator from being zero; its value is on a much smaller order of magnitude than the square of the normal operating current. (Function) Used to apply non-negative physical constraints, when the molecular calculation result is negative due to sensor noise (i.e., the measured heat dissipation is greater than the theoretical total heat), the contact impedance is forced to be set to zero to avoid non-physical negative impedance output.

[0088] In addition, the system has a preset minimum calculated current threshold. The threshold is set based on the system's signal-to-noise ratio (SNR), and the specific setting principle is as follows: The theoretical Joule heating rate generated at this current should be greater than the resolution of the temperature sensor (e.g., 0.1°C). Typically, this is... Set to 10% to 20% of the rated current of the monitoring node. When the measured current... If the system determines that the current Joule heating effect is insufficient for the sensor to effectively capture, it will pause updates and retain the previous estimate. ; when At that time, the system calculates and updates according to the above formula. Through this step, the system dynamically decomposes node temperature changes into environmental factors, heat dissipation factors, and contact factors. Even when fan start-stop causes drastic temperature fluctuations, or when nearby equipment heats up, leading to an increase in ambient temperature, as long as the physical properties of the contact resistance itself remain unchanged, the calculated... The sequence will remain relatively constant, thus achieving effective decoupling of fault characteristics from operating condition disturbances.

[0089] In step S60, impedance evolution trend analysis and fault early warning are performed. This step extracts deterministic degradation trends from the calculated impedance values ​​containing random noise through statistical analysis of time-series data, and verifies the effectiveness of fault signals by combining the dynamic characteristics of the load current.

[0090] Step S60 specifically includes the following sub-steps: S61, Smoothing and filtering of the impedance sequence. The trend analysis module reads the real-time contact impedance sequence output from step S50. The original calculation sequence exhibits high-frequency random jitter due to current sampling noise and differential calculation errors. The trend analysis module has a length of [length missing]. A sliding window is used to perform a moving average filter on the data within the window, and the output is a smoothed impedance value. .

[0091] The formula for smoothing filtering is: ; in, The value of depends on the system's sampling frequency and thermal time constant, for example, it can be set to the number of sampling points covering 1 to 5 minutes of data. This processing suppresses zero-mean random observation noise and preserves the low-frequency trend components of impedance changes.

[0092] S62, Calculation of relative drift. The system has a preset reference contact impedance. This benchmark value is determined during the stable operating period (e.g., the first 24 hours of operation) after the initial commissioning or maintenance reset of the electrical equipment. The impedance drift rate is obtained by performing arithmetic mean statistics and the trend analysis module calculates in real time the current smoothed impedance value relative to the reference value. : ; This drift rate quantifies the degree of degradation in the contact performance of the monitoring node relative to its initial state.

[0093] S63, Multidimensional Validation of Fault Characteristics. To distinguish between actual physical contact degradation and false alarms caused by model parameter errors (such as deviations in heat capacity estimation), the trend analysis module performs verification based on monotonicity and current independence.

[0094] First, perform monotonicity trend verification. The trend analysis module performs a window of recent time. Drift rate sequence (e.g., over the past hour) Perform linear regression analysis and calculate its slope. Only when When the value is positive and significantly greater than the zero-drift noise level, it is determined that there is an increasing impedance trend.

[0095] Secondly, a current-impedance decoupling verification was performed. This verification is based on the following physical principle: Contact resistance is a physical structural property of conductive contact surfaces (depending on the number of contact spots and the thickness of the oxide film), and macroscopically it should be independent of the magnitude of the load current flowing through it. The spurious impedance fluctuations caused by thermal model errors are usually strongly correlated with the load current (i.e., the larger the current, the larger the calculated impedance).

[0096] The trend analysis module calculates the smoothing impedance within the current sliding window. The relative change amplitude and load current The ratio of the relative magnitudes of change generates the coupling factor. : ; in, This represents the range (maximum minus minimum) or standard deviation of the data within the sliding window. and These are the mean values ​​within the window.

[0097] To avoid mathematical singularities, this calculation is performed only on the relative fluctuation range of the load current. If the current fluctuation exceeds the preset excitation threshold, it will be executed. If the current fluctuation is lower than the threshold, it is considered that the current is in a steady-state current condition and the decoupling cannot be verified by dynamic characteristics. The system will default to determining that the verification is passed or maintain the judgment result of the previous moment.

[0098] System setting decoupling threshold (For example, set to 0.1).

[0099] when When the impedance value changes independently of the current change, the calculated impedance value is deemed to have physical validity.

[0100] Conversely, if This indicates that the calculated impedance value changes with current fluctuations, and is therefore identified as a false signal and shielded.

[0101] S64, graded early warning generation, when the above verification passes and the impedance drift rate... When the preset threshold is exceeded, the trend analysis module outputs a warning message corresponding to the level.

[0102] The warning levels are set as follows: When At that time, a Level 1 warning signal is generated.

[0103] This threshold corresponds to the typical impedance increment range of the initial oxidation of the contact surface, prompting maintenance personnel to check the tightness of the node during the next inspection. when When the threshold is reached, a level 2 warning signal is generated. This threshold corresponds to the critical point where the contact spots decrease sharply and the system is about to enter the nonlinear temperature rise stage of thermal runaway. This indicates that there is a serious risk of poor contact at the node, and it is recommended to arrange an immediate shutdown for maintenance.

[0104] The warning information is uploaded to the host computer monitoring system or mobile terminal through the communication interface of the edge computing gateway. The information includes the fault node ID, current impedance value, drift rate and suggested measures.

[0105] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fault early warning method for electrical control cabinets based on deep learning, characterized in that, Includes the following steps: Acquire physical sensor data and control logic data within the electrical control cabinet, and perform timing alignment of the physical sensor data and control logic data based on a unified clock source; The control logic data is parsed to identify the current operating mode of the system, which includes a self-calibration mode and a monitoring mode. In the self-calibration mode, the ratio of the actual temperature drop rate to the theoretical temperature drop rate of the monitoring node is calculated, and the heat dissipation correction coefficient is updated. The heat dissipation correction coefficient characterizes the degree of attenuation of the actual heat dissipation capacity of the electrical control cabinet relative to the design value. In the monitoring mode, the control logic data is input into a preset depth parameter estimation network, which outputs nominal thermodynamic parameters and uses the heat dissipation correction coefficient to correct the nominal thermodynamic parameters to obtain effective thermodynamic parameters. A thermal balance differential equation for the monitoring node is constructed. The physical sensing data and the effective thermodynamic parameters are substituted into the inverse form of the thermal balance differential equation to calculate the real-time contact impedance of the monitoring node. The system monitors the time evolution trend of the real-time contact impedance and generates a fault warning signal when the real-time contact impedance shows a monotonically increasing trend and is decoupled from the change trend of the load current.

2. The method and system for early warning of electrical control cabinet faults based on deep learning according to claim 1, characterized in that, The physical sensing data includes the current value of the circuit where the monitoring node is located, the node temperature value of the monitoring node, and the ambient temperature value inside the electrical control cabinet. The control logic data includes start / stop commands for the cooling fan and load commands for nearby devices; The current operating mode of the identification system specifically includes: If the control logic data indicates that the cooling fan is on and the current value is lower than the preset zero drift threshold, it is determined to be in self-calibration mode. If the current value is greater than the zero drift threshold, it is determined to be the monitoring mode.

3. The method and system for early warning of electrical control cabinet faults based on deep learning according to claim 1, characterized in that, The updated heat dissipation correction factor specifically includes: During the cooling phase after the load current is cut off, the actual temperature drop rate of the monitoring node is collected. Based on Newton's law of cooling, the theoretical temperature drop rate is calculated using the current node temperature, ambient temperature, and reference heat transfer coefficient under design conditions. The ratio of the actual temperature drop rate to the theoretical temperature drop rate is determined as the current heat dissipation correction factor, and this factor is stored in the parameter holder for use by the monitoring mode.

4. The method and system for early warning of electrical control cabinet faults based on deep learning according to claim 1, characterized in that, The deep parameter estimation network is a pre-trained multi-layer neural network, and its input-output relationship is as follows: The input feature vector of the depth parameter estimation network includes at least: Fan operating status, ambient temperature, and overall external thermal characteristics; The comprehensive external thermal potential characteristics are calculated by weighting the load command of the nearby device, the geometric distance of the nearby device relative to the monitoring node, and the rated power of the nearby device. The output of the depth parameter estimation network is the nominal thermodynamic parameter, specifically including the nominal convective heat transfer coefficient and the external disturbance heat flux. The nominal convective heat transfer coefficient characterizes the theoretical heat transfer capacity under ideal clean conditions.

5. The method and system for early warning of electrical control cabinet faults based on deep learning according to claim 4, characterized in that, The correction of the nominal thermodynamic parameters using the heat dissipation correction coefficient specifically includes: Read the thermal correction factor updated in the most recent self-calibration mode; Multiply the nominal convective heat transfer coefficient by the heat dissipation correction coefficient to obtain the effective convective heat transfer coefficient; The external disturbance heat flux is directly used as the effective external heat flux; The effective thermodynamic parameters consist of the effective convective heat transfer coefficient and the effective external heat flux.

6. The method and system for early warning of electrical control cabinet faults based on deep learning according to claim 5, characterized in that, The thermal balance differential equation describes that the increase in internal energy of the monitoring node per unit time is equal to the algebraic sum of the Joule heat generated by the current, the effective external heat flux, and the convective heat dissipation power. The real-time contact impedance of the monitoring node is calculated by performing the following calculations: Calculate the thermal inertia term, which is the product of the equivalent heat capacity of the monitoring node and the node temperature change rate; Calculate the convective heat dissipation term, which is the product of the effective convective heat transfer coefficient, the effective heat dissipation area, and the node temperature rise; Add the thermal inertia term to the convective heat dissipation term, and subtract the effective external heat flux to obtain the total heat generation power; The real-time contact impedance is obtained by dividing the total heat generation power by the square of the current value.

7. The method and system for early warning of electrical control cabinet faults based on deep learning according to claim 6, characterized in that, The calculation of the real-time contact impedance of the monitoring node also includes a numerical stability processing step: A regularization parameter is introduced into the denominator of the division operation to prevent computational singularities when the current crosses zero. A non-negative physical constraint is introduced, which forces the real-time contact impedance to be set to zero when the calculated numerator term is less than zero. A minimum calculated current threshold is set, and the real-time contact impedance is only updated when the theoretical temperature rise rate generated by the current value is greater than the resolution of the temperature sensor; otherwise, the calculated value from the previous moment is retained.

8. The method and system for early warning of electrical control cabinet faults based on deep learning according to claim 1, characterized in that, The early warning model, decoupled from the load current change trend, specifically includes: A sliding window is established, and the real-time contact impedance and load current within the window are statistically analyzed. The coupling factor is obtained by calculating the ratio of the relative rate of change of the real-time contact impedance to the relative rate of change of the load current. When the coupling factor is less than the preset decoupling threshold, it is determined that the change in the real-time contact impedance is independent of the change in the load current, thus confirming the physical validity of the impedance data. When the coupling factor is greater than or equal to the decoupling threshold, it is determined that the impedance data is affected by model error, and the current warning signal is blocked.

9. The method and system for early warning of electrical control cabinet faults based on deep learning according to claim 8, characterized in that, The generation of the fault warning signal specifically includes: Obtain the reference contact impedance determined during the system initialization phase; Calculate the percentage drift of the current real-time contact impedance relative to the reference contact impedance; When the drift percentage is in the first range, a first-level warning signal is generated to prompt checking the fastening status; When the drift percentage is in the second range, a secondary warning signal is generated to prompt shutdown and maintenance; the lower limit of the second range is greater than the upper limit of the first range.

10. A deep learning-based electrical control cabinet fault early warning system, comprising the deep learning-based electrical control cabinet fault early warning method according to any one of claims 1-9, characterized in that, include: The data synchronization module is used to acquire physical sensor data and control logic data and perform timing alignment. The state mapping module is used to identify whether the system is in self-calibration mode or monitoring mode based on the control logic data. The parameter calibration module is used to update the heat dissipation correction coefficient based on the actual temperature drop rate in self-calibration mode. The parameter reconstruction module is used to generate effective thermodynamic parameters in monitoring mode using a deep parameter estimation network and the heat dissipation correction coefficient. The impedance calculation module is used to calculate the real-time contact impedance based on the physical sensing data and the effective thermodynamic parameters using a physical reverse decoupling algorithm. The trend analysis module is used to monitor the evolution trend of the real-time contact impedance and perform fault early warning.

Images