Intelligent monitoring system for relay protection hardboard state

Abstract

The application relates to the technical field of power system monitoring, and discloses a relay protection hard press plate state intelligent monitoring system. A state acquisition module of the system acquires the mechanical displacement amount, the contact resistance value and the surface infrared thermal imaging data of the hard press plate in real time; a feature analysis module processes the above data respectively to generate a displacement feature vector, a resistance feature matrix and a thermodynamic feature spectrum; a joint representation module inputs the features into a cascaded attention fusion network, and outputs a multi-source joint state representation through spatial alignment and weight distribution; an abnormality prediction module inputs the multi-source joint state representation into a lightweight time series prediction model embedded with contact physical constraints to predict the displacement offset amount, the resistance attenuation rate and the thermal spot diffusion trend of the hard press plate in the next three working periods; and a decision execution module generates a hard press plate maintenance instruction set containing a mechanical lubrication level, a contact polishing frequency and a heat dissipation fan rotating speed adjustment amount according to the prediction result.

Description

Technical Field

[0001] This invention relates to the field of power system monitoring technology, specifically to an intelligent monitoring system for the status of relay protection hard-plate. Background Technology

[0002] In modern society, the power system is like the "main artery" of society, the cornerstone of maintaining normal social operations. From lighting and household appliances in daily life, to the operation of large machinery in industrial production, to the operation of shopping malls and data centers in the commercial sector, and even the operation of electric vehicles and urban rail transit in the transportation system, electricity is ubiquitous. Once the power supply is disrupted, the entire society will fall into chaos, with inconvenience in daily life, stagnation of industrial production, interruption of commercial activities, and paralysis of the transportation system leading to incalculable economic losses and social impacts.

[0003] Relay protection, acting as the "intelligent guardian" of the power system, plays a crucial role. Like the power system's "immune system," it monitors the system's operational status in real time. When faults occur, such as short circuits, grounding faults, overcurrents, or overvoltages, relay protection devices react quickly, rapidly isolating the faulty portion to prevent its spread and ensuring the normal operation of other parts of the power system. This maintains the stability of the power system, protects electrical equipment from damage, and ensures the reliability and security of power supply, playing a vital protective role in the safe and stable operation of the entire power system.

[0004] In relay protection devices, the hard switch plate is an indispensable key component. It is typically installed on the secondary control circuit and is a mechanical structure that controls the connection or disconnection of the relay protection device. Manually engaging or disengaging the hard switch plate allows the relay protection device to be turned on or off, much like setting a "switch" for the device. This "switch" directly determines whether the relay protection device can function properly to protect the power system. For example, during power system maintenance, it may be necessary to deactivate certain relay protection devices; this can be achieved by disengaging the corresponding hard switch plate. During normal operation, the hard switch plate is engaged, placing the relay protection device in a standby state, ready to protect against potential faults. Therefore, the correct engagement or disengagement of the hard switch plate directly affects the effective implementation of the power system's protection functions and is a crucial link in ensuring the safe operation of the power system.

[0005] Traditionally, monitoring the status of rigid pressure plates relies primarily on manual inspections. Maintenance personnel need to periodically visit substations and other locations with stringent requirements for lighting conditions and equipment layout to meticulously inspect each of the numerous rigid pressure plates on the protection devices. The dense arrangement, large number, and generally small size of these pressure plates make manual inspection extremely difficult. During prolonged and intensive inspections, maintenance personnel are prone to visual fatigue, leading to inaccurate judgments of the pressure plate's engagement / disengagement status. Furthermore, memory lapses can cause recording errors, making it difficult to guarantee 100% accuracy in identifying incorrectly engaged or disengaged pressure plates.

[0006] When adjusting the operation of power grids, adjusting protection devices, or carrying out maintenance work that requires changing the switching status of relays, manual operation has many shortcomings. The operation process often involves long periods of time, lacks effective recording methods, makes it difficult to trace the operation process and time, and lacks effective error prevention measures. This means that the hidden danger of incorrect switching status of protection devices may persist for a long time, difficult to detect and correct in a timely manner, planting a "time bomb" for the safe operation of power equipment. Once an abnormal switching status triggers a relay protection device to malfunction or fail to operate during equipment operation, it will pose a serious threat to the safe and stable operation of the power system. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent monitoring system for the status of relay protection hard plate, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides an intelligent monitoring system for the status of relay protection hard-plate circuit boards, the system comprising:

[0009] The status acquisition module is used to acquire the mechanical displacement, contact resistance value and surface infrared thermal imaging data of the relay protection hard plate in real time.

[0010] The feature parsing module is used to perform dynamic trajectory decomposition of mechanical displacement using a parallel sparse coding network to generate displacement feature vectors. At the same time, it extracts the gradient distribution pattern of contact resistance values ​​through a multi-scale residual convolutional network to generate a resistance feature matrix. Based on an adaptive threshold segmentation algorithm, it performs contour marking on high-temperature areas in infrared thermal imaging data to generate a thermodynamic feature map.

[0011] The joint characterization module is used to input displacement feature vectors, resistance feature matrices and thermodynamic feature maps into a cascaded attention fusion network, correct feature dimension differences through a spatial alignment mechanism, and introduce learnable weight allocation coefficients in the channel dimension to output a multi-source joint state characterization.

[0012] The anomaly prediction module is used to embed the multi-source joint state characterization into a lightweight time-series prediction model of contact physical constraints, and output the predicted values ​​of displacement offset, resistance attenuation rate and hot spot diffusion trend in the next three working cycles.

[0013] The decision execution module is used to generate a set of hard plate maintenance instructions based on the predicted values ​​of displacement offset, resistance attenuation rate, and hot spot diffusion trend. The instruction set includes mechanical lubrication level, contact grinding frequency, and cooling fan speed adjustment.

[0014] Preferably, when the feature parsing module uses a parallel sparse coding network to perform dynamic trajectory decomposition of the mechanical displacement:

[0015] The time series of mechanical displacement is divided into overlapping sliding windows. The data in each window is learned by basis function through three sparse coding layers. The first layer extracts millisecond-level vibration components, the second layer separates second-level stroke fluctuation components, and the third layer fuses the outputs of the first two layers to generate a displacement feature vector.

[0016] When processing contact resistance values ​​using a multi-scale residual convolutional network, five sets of convolutional kernels with different widths are used to scan the resistance value sequence in parallel. The output of each set of convolutional kernels is compressed by depth-separable convolution and then concatenated with the original gradient features passed by skip connections to form a resistance feature matrix.

[0017] Preferably, when performing contour marking on high-temperature regions in infrared thermal imaging data based on the adaptive threshold segmentation algorithm:

[0018] First, bilateral filtering is used to eliminate thermal imaging noise. Then, the local entropy value of the image is calculated to dynamically adjust the segmentation threshold. For pixel clusters that exceed the threshold, morphological closing operation is performed to connect the broken edges. Finally, the closed polygon contour is extracted by the edge tracking algorithm and the centroid coordinates are marked to form a thermodynamic feature map.

[0019] Preferably, the operation of the cascaded attention fusion network in the joint representation module includes:

[0020] The displacement feature vector is spatially stacked with the resistance feature matrix after being transformed from one-dimensional to two-dimensional, and the resolution is adjusted by hollow spatial pyramid pooling.

[0021] A region proposal network is used to generate regions of interest from the thermodynamic feature map and map them to the same spatial scale as the displacement-resistance hybrid feature.

[0022] The contribution weights of the three types of features are dynamically calculated using gated cyclic units in the channel dimension, and the final output is a multi-source joint state representation that includes spatial alignment features.

[0023] Preferably, the lightweight time-series prediction model in the anomaly prediction module is constructed as follows:

[0024] In the model encoder, grouped convolutions are used to reduce the number of parameters, and long short-term memory skip connections are introduced in the decoder.

[0025] The physical constraints of the contact are transformed into regularization terms in the model loss function. The constraints include the inverse relationship between contact pressure and resistance, and the linear correlation between displacement and heat dissipation area.

[0026] The model output layer is connected in parallel to three regression heads, which correspond to the predicted values ​​of displacement offset, resistivity attenuation rate, and hot spot diffusion trend, respectively.

[0027] Preferably, the process by which the decision execution module generates the hard plate maintenance instruction set includes:

[0028] A decision tree containing twenty-four fault modes was established, with each branch node corresponding to different combinations of predicted values ​​for displacement offset, resistance attenuation rate, and hot spot diffusion trend.

[0029] Preset maintenance parameters are stored in the leaf nodes. When the predicted value combination falls within the coverage area of ​​a certain leaf node, the corresponding mechanical lubrication level, contact polishing frequency and cooling fan speed adjustment command is triggered.

[0030] Preferably, when acquiring surface infrared thermal imaging data in the status acquisition module:

[0031] A dual-band infrared sensor is used to simultaneously acquire long-wave and mid-wave infrared images, and a non-uniformity correction algorithm is used to eliminate detector response differences.

[0032] Pixel-level fusion of dual-band images generates composite thermal imaging data with temperature sensitivity gradients;

[0033] During the mechanical displacement acquisition process, a six-axis inertial measurement unit and a laser displacement sensor are integrated to record the three-dimensional motion trajectory of the hard platen with millimeter-level accuracy.

[0034] Preferably, the structure of the multi-scale residual convolutional network in the feature parsing module includes:

[0035] The first scale uses a convolution kernel with a width of 5 to capture the macroscopic trend changes in resistance values;

[0036] The second scale uses a convolutional kernel with a width of 3 to extract local mutation features;

[0037] The third scale expands the receptive field through dilated convolution with a dilation rate of 2;

[0038] After being processed by the batch normalization layer, the outputs at each scale are fused element-wise with the feature maps passed through the cross-scale jump connections.

[0039] Preferably, the spatial alignment mechanism in the joint representation module is implemented as follows:

[0040] The displacement eigenvectors are extended to a two-dimensional mesh using cubic spline interpolation.

[0041] Fill the blank areas of the resistance feature matrix with Laplace diffusion results based on neighboring resistance values;

[0042] By adjusting the geometric deformation of the thermodynamic feature map using deformable convolutional networks, the key regions are made consistent with the spatial distribution of displacement-resistance features.

[0043] Preferably, the process of constructing the fault mode decision tree in the decision execution module includes:

[0044] Five thousand sets of predicted values ​​and actual maintenance parameters were collected from historical maintenance records as training samples.

[0045] The Gini coefficient is used as the splitting criterion for recursive feature partitioning.

[0046] After eliminating overfitting branches using a pruning algorithm, each leaf node is bound to the standard operating procedure in the equipment maintenance manual.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] The status acquisition module of this invention possesses powerful real-time monitoring capabilities, enabling comprehensive and uninterrupted acquisition of mechanical displacement, contact resistance values, and surface infrared thermal imaging data of the relay protection hard plate. This innovative design allows the system to track the operating status of the hard plate in real time, completely changing the lag and incompleteness of previous manual inspections. In traditional manual inspection modes, due to time and energy constraints, maintenance personnel find it difficult to monitor the hard plate's status in real time, often only able to conduct inspections within the prescribed inspection cycle. The real-time monitoring function of this system is like installing a pair of "24-hour eyes" on the hard plate, capable of promptly capturing any subtle changes in its status. Whether day or night, whether in normal operation or during grid operation adjustments, protection device adjustments, or maintenance work, the system operates stably, ensuring that no factor that might affect the normal operation of the hard plate is missed, providing comprehensive and timely data support for subsequent analysis and decision-making.

[0049] The feature analysis module is a major technical highlight of this invention. It employs advanced technologies such as parallel sparse coding networks, multi-scale residual convolutional networks, and adaptive threshold segmentation algorithms to perform in-depth and detailed analysis of the data acquired by the state acquisition module. For mechanical displacement, the parallel sparse coding network can accurately decompose its complex dynamic trajectory, generating a representative displacement feature vector. This vector acts like a "digital fingerprint" of the mechanical displacement state, accurately reflecting the displacement change trend of the rigid plate at different times. Simultaneously, the multi-scale residual convolutional network can keenly capture the gradient distribution pattern in the contact resistance value, generating a resistance feature matrix. This matrix records in detail the changes in contact resistance across different dimensions, helping to uncover potential problems hidden behind resistance changes. Based on the adaptive threshold segmentation algorithm, high-temperature areas in the infrared thermal imaging data are outlined, and the generated thermodynamic feature map intuitively displays the temperature distribution on the surface of the rigid plate, making thermal anomalies readily apparent. Compared to traditional data analysis methods, the feature analysis module of this invention can more deeply and accurately mine key information in the data, providing a more reliable and accurate basis for subsequent judgment of the rigid plate's state.

[0050] The joint characterization module plays a crucial role as the "data fusion brain" in this invention, organically fusing information from three different data sources: displacement feature vectors, resistance feature matrices, and thermodynamic feature spectra. During the fusion process, a spatial alignment mechanism cleverly corrects dimensional differences between features, much like precisely assembling jigsaw puzzle pieces of varying sizes for seamless integration. Simultaneously, a learnable weighting coefficient is introduced into the channel dimension. This design is akin to assigning different "importance labels" to different information, allocating appropriate weights to each feature based on its importance and relevance, thereby highlighting key information and downplaying secondary information. Through this processing, the output multi-source joint state characterization comprehensively and accurately reflects the actual operating state of the hard platen. This multi-source information fusion approach avoids the bias and limitations that may arise from a single data source, enabling the monitoring system to gain a deeper and more comprehensive understanding of the hard platen's state, providing more comprehensive and reliable foundational data for subsequent anomaly prediction and decision-making. Attached Figure Description

[0051] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent monitoring system for the status of the relay protection hard plate described in this invention.

[0052] Figure 2 This is a flowchart illustrating the working principle of high-temperature region contour marking in infrared thermal imaging data based on an adaptive threshold segmentation algorithm.

[0053] Figure 3This is a flowchart illustrating the working principle of the cascaded attention fusion network operation in the joint representation module. Detailed Implementation

[0054] 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.

[0055] Please see Figure 1 This invention provides an intelligent monitoring system for the status of relay protection hard-plate circuit boards. The system comprises a status acquisition module, a feature analysis module, a joint characterization module, an anomaly prediction module, and a decision execution module operating collaboratively. Specific implementation methods are as follows:

[0056] The status acquisition module acquires real-time mechanical displacement, contact resistance, and surface infrared thermal imaging data of the relay protection hard plate. Mechanical displacement is recorded by a high-precision sensor, and contact resistance is measured using Ohm's law to determine circuit continuity. Surface infrared thermal imaging captures temperature distribution using an infrared camera. The feature parsing module uses a parallel sparse coding network to dynamically decompose the mechanical displacement, generating a displacement feature vector. Simultaneously, a multi-scale residual convolutional network extracts the gradient distribution pattern of the contact resistance, generating a resistance feature matrix. An adaptive threshold segmentation algorithm is used to outline high-temperature regions in the infrared thermal imaging data, generating a thermodynamic feature map. The joint characterization module inputs the displacement feature vector, resistance feature matrix, and thermodynamic feature map into a cascaded attention fusion network. A spatial alignment mechanism corrects for feature dimension differences, and learnable weight allocation coefficients are introduced in the channel dimension to output a multi-source joint state characterization. The anomaly prediction module inputs the multi-source joint state characterization into a lightweight time-series prediction model embedded with contact physical constraints, outputting predicted values ​​for displacement offset, resistance attenuation rate, and hot spot diffusion trend over the next three operating cycles. The decision execution module generates a set of maintenance instructions for the hard plate based on the predicted values ​​of displacement offset, resistance attenuation rate, and hot spot diffusion trend. The instruction set includes mechanical lubrication level, contact grinding frequency, and cooling fan speed adjustment, thereby realizing automated maintenance decision-making.

[0057] Example 1: See Figure 2In the feature parsing module, when using a parallel sparse coding network to dynamically decompose the mechanical displacement, the time series of the mechanical displacement is divided into overlapping sliding windows. The overlapping sliding window setting ensures the continuity of the time series and avoids boundary effects. The data within each window undergoes basis function learning through three sparse coding layers. The basis function learning process uses an iterative optimization algorithm to minimize the reconstruction error. The first sparse coding layer extracts millisecond-level vibration components, which reflect the high-frequency mechanical vibration signals during the operation of the relay protection hard plate. The vibration signals include the impact response characteristics at the moment of contact closure. The second sparse coding layer separates second-level stroke fluctuation components, which capture the trajectory patterns formed by the low-speed movement of the hard plate operating mechanism. The trajectory patterns reflect the wear accumulation effect of the mechanical transmission components. The third sparse coding layer fuses the outputs of the first two layers and generates a displacement feature vector. The fusion operation retains multi-timescale feature information through weighted concatenation.

[0058] In processing contact resistance values ​​using a multi-scale residual convolutional network, five sets of convolutional kernels with different widths are used to scan the resistance value sequence in parallel. The variation in kernel width covers both macroscopic trends and local details. The output of each set of kernels is compressed using depthwise separable convolution, which decomposes the standard convolution into depthwise convolution and pointwise convolution to reduce computational complexity. The compressed features are concatenated with the original gradient features passed through skip connections to form a resistance feature matrix. Skip connections preserve the original gradient information to avoid feature loss. The structure of the multi-scale residual convolutional network includes: a first scale using a 5-width convolutional kernel to capture macroscopic trend changes in resistance values, reflecting the long-term resistance drift caused by aging of the contact material; a second scale using a 3-width convolutional kernel to extract local abrupt changes, identifying instantaneous resistance fluctuations caused by arc erosion on the contact surface; and a third scale expanding the receptive field through dilated convolution with an expansion rate of 2, enhancing context awareness while maintaining parameter quantity. The outputs of each scale are processed by batch normalization layers, which accelerate network training convergence and stabilize gradient propagation. The processed features are fused element-wise with the feature maps passed by the cross-scale skip connections. The cross-scale skip connections promote the reuse and complementarity of multi-scale features.

[0059] When using an adaptive threshold segmentation algorithm to outline high-temperature regions in infrared thermal imaging data, bilateral filtering is employed to eliminate thermal imaging noise. Bilateral filtering removes random noise while preserving edge details of the hotspot region. The segmentation threshold is dynamically adjusted by calculating the local entropy value of the image, quantifying the complexity of local image regions to adapt to different thermal distribution scenarios. Morphological closing operations are performed on pixel clusters exceeding the threshold to connect broken edges. These morphological closing operations fill internal voids in the hotspot region through a dilation-erosion process. Closed polygonal contours are extracted and centroid coordinates are labeled using an edge tracking algorithm, which extracts complete hotspot boundaries based on gradient direction continuity. The resulting thermodynamic feature map contains a spatial mapping relationship between the hotspot geometry and temperature distribution. During mechanical displacement feature analysis, the parallel sparse coding network is trained using the K-SVD dictionary learning algorithm, which constructs an overcomplete set of basis functions based on sparse representation theory. The step size of the overlapping sliding window is set to one-third of the window length, balancing the timeliness and computational efficiency of feature extraction. Millisecond-level vibration components are extracted using a high-frequency bandpass filter preprocessing, with the bandpass filter cutoff frequency set according to the mechanical resonance characteristics of the hard-plate. The separation of second-level travel fluctuation components is aided by empirical mode decomposition (EMD), which decomposes non-stationary signals into essential mode functions. The dimensionality of the displacement feature vector is reduced using principal component analysis (PCA), which retains 95% of the variance contribution rate of the original features.

[0060] In the multi-scale feature extraction of contact resistance values, the widths of the five convolutional kernels are set to 5, 3, 7, 9, and 11, respectively, covering the complete feature spectrum from instantaneous abrupt changes to long-term drift. The channel compression ratio of the depthwise separable convolution is 4:1, balancing feature representation capability and computational resource consumption. A 1×1 convolution is added to the skip connection path for dimension matching, allowing for flexible adjustment of the number of feature channels. The momentum parameter of the batch normalization layer is set to 0.9, controlling the update speed of the moving average statistics. Cross-scale skip connections employ a feature pyramid structure, enabling top-down feature enhancement propagation. The bilateral filter spatial standard deviation in the infrared thermal imaging processing is set to 3 pixels, and the range standard deviation is set to 0.1, optimized based on the noise characteristics of the thermal imaging sensor. Local entropy calculation uses a 7×7 sliding window, ensuring the stability and sensitivity of local statistics. Morphological closing operations use circular structuring elements with a radius of 5 pixels, maintaining the isotropic processing of the hot spot contour. The edge tracking algorithm uses the Canny operator combined with contour topology analysis, with the Canny operator's dual thresholds set to 20% and 40% of the maximum gradient of the image. The coordinate system of the thermodynamic feature map is rigidly transformed and registered with the spatial coordinate system of the mechanical displacement acquisition space, and the coordinate registration achieves spatial consistency alignment of multimodal data.

[0061] The feature parsing module employs a multi-core processor parallel computing architecture, with independent computing units allocated to process mechanical displacement, contact resistance, and infrared thermal imaging data. Mechanical displacement parsing is handled by a digital signal processing (DSP) core, which optimizes numerical computation efficiency. Contact resistance processing is handled by a graphics processing (GPU) unit, which accelerates the convolutional neural network inference process. Infrared thermal imaging processing is handled by a general-purpose computing core, which processes image analysis algorithms. Internal data exchange utilizes a double-buffered pipeline mechanism, ensuring continuous data acquisition and processing. The real-time requirement for dynamic trajectory decomposition of mechanical displacement is met through a sliding window pipeline, which overlaps data segmentation, feature extraction, and vector generation operations. Multi-scale scanning of contact resistance is implemented using a convolutional computation optimization library, which leverages the SIMD instruction set to process multi-scale convolutional kernels in parallel. Adaptive threshold segmentation for infrared thermal imaging employs multi-threaded parallel region processing, dividing the thermal imaging image into multiple sub-regions for simultaneous computation. The feature parsing module output interface adopts a standardized data encapsulation format, which includes feature dimensions, timestamps, and data verification information.

[0062] The number of basis functions in the parallel sparse coding network is adaptively adjusted based on the amount of training data, maintaining a logarithmic ratio with the number of training samples. The multi-scale residual convolutional network is trained using a momentum-driven stochastic gradient descent algorithm, with a momentum parameter set to 0.9 to accelerate convergence. The adaptive threshold segmentation algorithm employs an online learning strategy for parameter updates, dynamically optimizing threshold calculation parameters based on the latest thermal imaging data. The feature parsing module's power management utilizes dynamic voltage and frequency adjustment technology, dynamically adjusting the processor's operating state in real-time according to the computational load. The anti-interference capability of mechanical displacement feature parsing is enhanced by adding noise injection data, simulating signal distortion under on-site vibration conditions. The robustness of contact resistance feature extraction is enhanced through multi-sensor data fusion, eliminating measurement errors from individual sensors. The accuracy of infrared thermal imaging contour marking is improved with the assistance of multispectral information, which provides visible light images to aid edge verification. The feature parsing module's reliability design employs a redundant computing unit backup mechanism, automatically switching to the backup computing unit in case of hardware failure.

[0063] Example 2: See Figure 3The operations of the cascaded attention fusion network in the joint representation module include: converting the displacement feature vector from one dimension to two dimensions and then spatially stacking it with the resistance feature matrix; the one-dimensional to two-dimensional conversion operation reconstructs the time-series data of mechanical displacement into a two-dimensional grayscale image format through temporal padding, where the grayscale value of each pixel corresponds to the displacement value at a specific time; the spatial stacking operation concatenates the two-dimensional image converted from the displacement feature vector with the resistance feature matrix along the channel dimension, forming a preliminary fusion tensor of multimodal features; the resolution is adjusted by dilated spatial pyramid pooling, which uses four parallel convolutional branches with different dilation rates to extract multi-scale contextual information, with the convolutional kernel size fixed at 3×3 and the dilation rates set to 1, 2, 4, and 8 respectively; a region proposal network is used to generate regions of interest on the thermodynamic feature map, which slides across the thermodynamic feature map based on an anchoring mechanism to generate candidate region bounding boxes and their feature vectors; these are mapped to the same spatial scale as the displacement-resistance hybrid features, and the mapping process uses bilinear interpolation to adjust the feature map size and affine transformation to achieve coordinate system unification.

[0064] In the channel dimension, gated recurrent units dynamically calculate the contribution weights of the three types of features. These units process displacement, resistance, and thermodynamic features sequentially, updating the retention ratio of historical information by the gate control and resetting the gate to determine the fusion degree of the current input. The final output is a multi-source joint state representation containing spatially aligned features. The dimension of this representation is compressed into a 256-dimensional floating-point vector, with each dimension containing the spatially aligned feature response value. The spatial alignment mechanism is implemented by extending the displacement feature vector to a two-dimensional grid using cubic spline interpolation. This cubic spline interpolation ensures a smooth and continuous surface on the two-dimensional grid, and the grid node values ​​maintain monotonically consistent with the original displacement feature vector. Blank areas in the resistance feature matrix are filled with Laplace diffusion results based on neighboring resistance values. The Laplace diffusion equation is solved iteratively to achieve a smooth transition of values ​​in the blank areas, with the diffusion coefficient inversely proportional to the resistance gradient. A deformable convolutional network adjusts the geometric deformation of the thermodynamic feature map. This network learns spatial transformation parameters to predict the offset of each convolutional kernel, with the offset constrained within the range of [-1, 1] to ensure transformation stability. To ensure that the key region is consistent with the spatial distribution of displacement-resistance characteristics, spatial distribution consistency is achieved by minimizing the feature mutual information loss function, and mutual information is calculated using Monte Carlo sampling estimation.

[0065] The cascaded attention fusion network comprises three stages: feature preprocessing, spatial alignment, and weight fusion. The feature preprocessing stage normalizes the input features and scales the values ​​to the [0,1] range. The spatial alignment stage combines a deformable convolutional network with an interpolation algorithm. The deformable convolutional network has three layers, each with 64 kernels. The weight fusion stage employs a cross-modal attention mechanism, where the query vector originates from displacement features, the key vector from resistance features, and the value vector from thermodynamic features. The output feature map from the hollow spatial pyramid pooling is dimensionality-reduced by 1×1 convolution and then residually connected to the original features. This residual connection avoids the gradient vanishing problem in deep networks. Candidate regions generated by the region proposal network are filtered using non-maximum suppression (NMS), with a NMS threshold of 0.7, retaining high-quality candidate boxes with overlap below the threshold. The hidden layer dimension of the gated recurrent unit is set to 512, and the hidden layer state initialization uses the Xavier uniform initialization method. The cubic spline interpolation in the spatial alignment mechanism ensures continuous boundary conditions, which are set as natural boundary conditions, i.e., the second derivative is zero. The Laplacian diffusion calculation uses the Jacobi iteration method, with 100 iterations and a convergence tolerance of 1E-6. Training the deformable convolutional network requires adding regularization constraints; the regularization term penalizes excessive deformation offsets to prevent mesh distortion.

[0066] The storage format for the multi-source joint state representation adopts tensor form, with tensor dimensions of [batch_size, 256, 16, 16], where 16×16 represents the spatial resolution. The representation vector contains positional encoding information, which uses sine and cosine functions to generate absolute positional information. The computational complexity of the cascaded attention fusion network is controlled by grouped convolution, which divides the channels into 8 groups for independent computation before merging. At the hardware implementation level, the cascaded attention fusion network is deployed on a processor with tensor computation cores, supporting FP16 half-precision floating-point operations. Memory allocation employs a double-buffering mechanism: one buffer stores the input features, and the other performs fusion computation. The data pipeline is designed as a three-stage pipeline, with feature preprocessing, spatial alignment, and weight fusion stages executed in parallel. Interpolation computation in the spatial alignment mechanism requires dedicated interpolator hardware support, implementing a cubic convolution interpolation algorithm. Laplacian diffusion computation requires a parallel processor array, with each processing unit responsible for diffusion computation in a sub-region of the image. Deformable convolutional networks require programmable logic devices to predict deformation parameters, and the deformation parameter prediction network adopts a lightweight fully connected layer structure.

[0067] The output interface of the multi-source joint state representation adopts a high-speed serial bus with a bus width of 256 bits and a clock frequency of 1GHz. Timing control uses multi-phase clock synchronization, with a clock phase difference of 90 degrees to ensure data setup and hold time. Power management uses dynamic voltage and frequency adjustment technology to adjust the supply voltage and clock frequency according to the network load. The error detection mechanism of the cascaded attention fusion network includes parity check codes, which cover feature data and control signals. Fault tolerance design adopts triple module redundancy, with three copies of the computational units on the critical path for voting decisions. Real-time performance is guaranteed through worst-case execution time analysis, and the time limit for each processing stage is determined through static analysis. Network parameter updates use the stochastic gradient descent algorithm, with a learning rate set to 0.01 and decaying by a factor of 0.1 every 10 cycles. The gradient pruning threshold is set to 1.0 to prevent gradient explosion. Training data augmentation uses random rotation and scaling, with rotation angles ranging from -15° to 15° and scaling ratios ranging from 0.8 to 1.2. The one-dimensional to two-dimensional transformation of displacement feature vectors needs to preserve temporal relationships. In the transformed two-dimensional image, the row direction represents the time axis, and the column direction represents the displacement value. Filling blank areas in the resistance feature matrix requires boundary processing; mirror filling is used to avoid edge effects. The region proposal network for thermodynamic feature maps requires multi-scale anchors, with anchor sizes set to 8×8, 16×16, and 32×32. The weight matrix of the gated recurrent unit needs orthogonal initialization to ensure the stability of gradient propagation. Attention weights are calculated using scaled dot product attention, with the scaling factor being the reciprocal of the square root of the feature dimension. Dimensionality compression of the multi-source joint state representation uses principal component analysis (PCA), which retains 95% of the original feature variance.

[0068] The cascaded attention fusion network uses a standardized data format for its interfaces with the preceding and following modules. Data packets include timestamps, sequence numbers, and cyclic redundancy check codes. Synchronization between modules employs a hardware semaphore mechanism, with semaphores controlling mutually exclusive access for data read and write. An automatic retransmission request mechanism is used for error recovery; erroneous data packets trigger a retransmission process. Numerical precision management during network computation utilizes mixed-precision training: FP16 precision for forward computation and FP32 precision for backpropagation. Memory access is optimized for sequential access, with feature data stored in memory in row-major order. The caching strategy uses the Least Recently Used (LRU) algorithm, with cache block sizes aligned to 4KB memory pages. Quality assessment of multi-source joint state representation uses a reconstruction error metric, calculated by the decoder network as the mean squared error between the original and reconstructed features. A real-time monitoring module monitors feature distribution shifts; shift detection uses KL divergence to calculate the difference between the current and historical feature distributions. An anomaly handling mechanism includes a feature recalculation process, where abnormal features trigger recalculation without interrupting system operation.

[0069] Example 3: The lightweight temporal prediction model in the anomaly prediction module is constructed by using grouped convolution in the encoder to reduce the number of parameters. Grouped convolution divides the input channel into multiple independent subgroups and performs convolution operations within each subgroup. The decoder introduces long short-term memory (LSTM) skip connections, which directly transmit the hidden states of different encoder levels to the corresponding decoder levels. Contact physical constraints are transformed into regularization terms in the model loss function, including the inverse relationship between contact pressure and resistance. The model output layer connects three regression heads in parallel, corresponding to the predicted values ​​of displacement offset, resistance attenuation rate, and hot spot diffusion trend, respectively.

[0070] The training process of the lightweight time series prediction model uses multi-source joint state representations as input data. These representations are standardized to eliminate differences in feature dimensions. The loss function is designed as a weighted sum of the prediction error term and the physical constraint regularization term. The weighting coefficients are determined to have an optimal balance point through grid search. The prediction error term calculates the mean square error between the predicted and actual values. The mathematical expression for the physical constraint regularization term is:

[0071]

[0072] in: This indicates the contact resistance value (unit: Ω). Indicates contact force (unit: N). A physical constant (unit: S / N) characterizing the proportional relationship between electrical conductance and force. is the regularization coefficient.

[0073] The model training employs a mini-batch gradient descent algorithm with a batch size of 32 samples. The initial learning rate is 0.001, adjusted using cosine annealing. The maximum number of gradient descent iterations is set to 1000, and an early stopping mechanism is triggered when the validation set loss fails to improve for 10 consecutive cycles. Model parameters are initialized using the He normal distribution, which is suitable for the ReLU activation function family and accelerates training convergence. During the inference phase of the lightweight temporal prediction model, real-time multi-source joint state representations are received. These representations are preprocessed and fed into the encoder's grouped convolutional layers, where the output feature maps are compressed to one-quarter of the input size. The decoder's long short-term memory skip connections receive features from each encoder layer and fuse them with time-series information, including the correlation patterns between historical and current states. Three regression heads output predicted value sequences for the next three working cycles, each sequence containing a continuous trend over time. The output format of the displacement offset prediction is a three-dimensional vector representing the X / Y / Z axial offset; the output of the resistance attenuation rate prediction is a scalar value in ohms per hour; and the output of the hot spot diffusion trend prediction is a sequence of coordinates of the hot spot centroid movement trajectory.

[0074] During model deployment, quantization techniques are employed to reduce memory consumption, quantizing weight parameters from FP32 precision to INT8 precision. The quantization process includes calibration steps to minimize accuracy loss. The inference engine is optimized to a single-instruction, multiple-data-stream architecture, processing computational tasks across multiple feature channels in parallel. Real-time requirements are met through layer fusion technology, which combines convolutional layers and batch normalization layers into a single computational unit. The lightweight temporal prediction model's online learning function supports dynamic parameter updates. Online learning uses a sliding window mechanism to retain the latest 1000 sets of sample data, and model parameters are retrained every 24 hours. Model version management employs a blue-green deployment strategy, ensuring uninterrupted system operation during updates. An anomaly detection mechanism monitors model prediction bias, triggering a model retraining process when the prediction bias exceeds a threshold.

[0075] The multi-source joint state representation of the model's input features requires validity verification. Validity verification checks the range and distribution patterns of feature values, and abnormal feature values ​​trigger a data re-acquisition process. The output predicted values ​​undergo post-processing smoothing filtering, which uses a moving average method to eliminate random fluctuations. Uncertainty quantification of prediction results employs the Monte Carlo dropout method, which randomly discards neurons during the inference phase to generate multiple prediction samples. The robustness of the lightweight time-series prediction model is enhanced through adversarial training, which adds small perturbations to the samples to improve model robustness. Cross-device generalization capability is achieved through domain adaptation technology, which aligns with differences in data distribution across different devices. Long-term prediction stability is maintained through periodic recalibration, which compares the cumulative error between predicted and actual measured values. Model computational resource allocation uses dynamic priority scheduling, adjusting the model's computation frequency based on the hard-plate's operating status. Memory management employs a garbage collection mechanism to periodically release intermediate computation results, with the garbage collection threshold set at 80% of memory usage. Power consumption control is achieved through dynamic voltage and frequency adjustment technology, which automatically adjusts the voltage and frequency according to the computational load. The lightweight time-series forecasting model and the decision execution module use an asynchronous communication model, which avoids blocking the system's main thread during model inference. Data serialization uses a protocol buffer format, which provides efficient data encoding and decoding performance. The error handling mechanism includes prediction timeout detection, with a timeout threshold set to 100 milliseconds to ensure real-time response.

[0076] The model monitoring system records prediction accuracy and inference latency metrics. Accuracy metrics include mean absolute error and root mean square error, while latency metrics record the complete processing time from input to output. The logging system uses a structured log format, which facilitates subsequent analysis and debugging. Performance analysis tools pinpoint computational bottlenecks through code profiling and hardware counter analysis. The maintenance process for the lightweight time series prediction model includes regular model evaluation, which uses independent test datasets to calculate performance metrics. Parameter tuning is based on evaluation results and includes optimization of learning rate, batch size, and network structure. A version rollback mechanism is activated when model performance degrades, restoring the model parameters to the previous stable version. Model security is protected using digital signature technology, which verifies the integrity and authenticity of model files. Access control is based on role-based access control, differentiating between system administrators, maintenance engineers, and monitoring operators. Data encryption uses the AES-256 algorithm to protect training data and model parameters, with the AES-256 encryption key rotated periodically to enhance security. The lightweight time series prediction model's scalability design supports parallel prediction across multiple hardboards, achieved through model instantiation and resource isolation. The load balancing algorithm distributes computational tasks across multiple processing cores, taking into account the current utilization of each core and the priority of the predicted tasks. Horizontal scaling is achieved through containerized deployment, which allows for the rapid replication of model instances to cope with load growth.

[0077] The ideal contact state constant C in the physical constraint regularization term characterizes the constant relationship that the product of contact resistance R and contact pressure P should maintain under ideal contact conditions. This relationship is derived from Holm contact theory in contact physics. Regularization coefficient Controlling the influence of physical constraints on the total loss function An excessively large value can cause the model to overfit physical laws and ignore data features. If the value is too small, it will not effectively guide the model to follow physical laws. Performance tuning on the validation set is performed during training. The model aims to achieve an optimal balance between data-driven approaches and physical laws. Preprocessing of the training data includes outlier removal and missing value imputation. Outlier removal uses the three-standard-deviation principle, and missing value imputation uses time-series interpolation methods. Feature engineering includes the construction of temporal features, such as sliding window statistics, difference features, and periodic features. Model validation employs time-series cross-validation to maintain the temporal order of the data and prevent future information leakage. The lightweight time-series prediction model's network structure consists of an encoder and a decoder. The encoder comprises six layers of grouped convolutions, each followed by a batch normalization layer and a ReLU activation function. The decoder consists of three layers of long short-term memory networks, with residual connections added between each layer to prevent gradient vanishing. The three regression heads in the output layer use the same network structure but are trained independently. Each regression head contains two fully connected layers and one output layer.

[0078] Example 4: The process of generating a hard platen maintenance instruction set by the decision execution module includes establishing a decision tree containing twenty-four fault modes. Each branch node of the decision tree corresponds to different combination intervals of predicted displacement offset, predicted resistance attenuation rate, and predicted hot spot diffusion trend. Preset maintenance parameters are stored in the leaf nodes. When a combination of predicted values ​​falls within the coverage interval of a certain leaf node, the corresponding instructions for adjusting the mechanical lubrication level, contact polishing frequency, and cooling fan speed are triggered. The depth of the decision tree is set to 6 levels. Each non-leaf node contains a feature splitting condition, which is a binary split based on a threshold of the predicted value. The construction process of the fault mode decision tree includes collecting 5,000 sets of predicted values ​​and actual maintenance parameter combinations from historical maintenance records as training samples. The training samples contain the correspondence between hard platen operation data and maintenance logs over the past five years. A recursive feature split is performed using the Gini coefficient as the splitting criterion. The Gini coefficient calculates the impurity of the node data, and the feature that maximizes the reduction of the Gini coefficient of the child node is selected for splitting. After eliminating overfitted branches through a pruning algorithm, the pruning algorithm uses cost complexity pruning, with the complexity parameter α set to 0.01 to balance the tree size and the goodness of fit. Each leaf node is linked to the standard operating procedure (SOP) in the equipment maintenance manual. The SOP clearly defines the classification criteria for mechanical lubrication levels, the method for determining the frequency of contact grinding, and the calculation formula for cooling fan speed adjustments. The decision tree's twenty-four fault modes correspond to different degradation states of the hard platen. Fault mode classification is based on three levels: normal / warning / dangerous (based on displacement offset prediction), slow / fast / rapid (based on resistance attenuation rate prediction), and local / expanding / diffuse (based on hot spot diffusion trend prediction). The specific values ​​of maintenance parameters for each leaf node are determined based on expert experience and historical data statistical analysis. The statistical analysis uses a clustering algorithm to classify samples with similar maintenance needs. The decision tree reasoning process uses a depth-first search algorithm, recursively traversing from the root node to the leaf nodes.

[0079] The real-time decision-making process of the decision execution module is executed every five minutes, and the latest output value of the anomaly prediction module is read in real-time. Predicted value preprocessing includes normalization and outlier filtering. Normalization maps predicted values ​​of different dimensions to the [0,1] interval. During the decision tree traversal, each node compares the predicted value with the threshold, and the comparison result determines whether to traverse the left or right branch. After reaching the leaf node, the stored maintenance parameter combination is retrieved and encapsulated in a structured data format. The maintenance instruction set includes control parameters in three dimensions: mechanical lubrication level, contact grinding frequency, and cooling fan speed adjustment. The mechanical lubrication level is divided into four levels: no lubrication, light lubrication, standard lubrication, and enhanced lubrication. The contact grinding frequency is expressed as the number of times it is ground per week, with a value range of [0,7]. The cooling fan speed adjustment is a percentage change relative to the base speed, with an adjustment range limited to [-20%, +30%]. The interface between the decision execution module and the actuator adopts the industrial Ethernet communication protocol. The industrial Ethernet frame format includes the target device address, instruction type, and parameter value. Command transmission uses a request-response mode with a request-response timeout of 3 seconds. The security mechanism includes a command checksum and a sequence number; the checksum uses the CRC32 algorithm to detect transmission errors.

[0080] The decision tree update mechanism supports both online learning and offline training modes. Online learning dynamically adjusts node thresholds based on newly generated maintenance records. Offline training periodically rebuilds the entire decision tree, with a training cycle set to once every three months. Version management records the time and content of each decision tree structure change. The fault tolerance handling of the decision execution module includes strategies for handling missing prediction values; when a prediction value is missing, the most recently available value is used as a substitute or a re-prediction is triggered. When a decision tree traversal anomaly occurs, a backup rule set is activated, which is based on simplified logical judgment of maintenance needs. The system log records the input values, traversal paths, and output instructions for each decision process. Refer to Table 1 for the specific values ​​of the maintenance instructions corresponding to the twenty-four leaf nodes of the fault mode decision tree.

[0081] Table 1: Correspondence between Failure Modes and Maintenance Parameters

[0082]

[0083] The Gini coefficient calculation for the decision tree splitting criterion involves probability distribution estimation, based on the frequency of each category in the training samples. The stopping conditions for recursive feature splitting include a node sample count of less than 50 or a Gini coefficient decrease of less than 0.01. The cost complexity calculation of the pruning algorithm comprehensively considers both the complexity of the tree structure and the classification accuracy. The 5,000 sets of training samples in the historical maintenance records contain real data from the field operating environment, and this real data has undergone data cleaning to remove obviously erroneous records. The standard operating procedures in the equipment maintenance manual are based on the equipment manufacturer's technical specifications, which clearly define the maintenance operation standards corresponding to various fault modes. The binding of decision tree leaf nodes to standard operating procedures is achieved through unique identifiers, which establish a mapping relationship during system initialization. The hardware platform of the decision execution module uses an industrial-grade programmable logic controller (PLC), which has both digital input / output and analog input / output modules. The mechanical lubrication level control output is a 4-20mA analog signal, the contact grinding frequency control is achieved through relay pulse signal output, and the cooling fan speed adjustment uses PWM modulation waveform control. The generation of maintenance instruction sets is triggered by two modes: time-triggered and event-triggered. Time-triggered instructions execute a decision-making process every five minutes, while event-triggered instructions initiate a decision immediately when a predicted value suddenly exceeds a threshold. An instruction priority mechanism handles situations where multiple trigger conditions are met simultaneously, with priority determined by the severity of the fault mode.

[0084] The human-machine interface of the decision execution module displays the current fault mode diagnosis results and maintenance instructions. The diagnosis results are displayed in color codes to indicate the severity of the fault. An operator confirmation mechanism requires that important maintenance instructions must be manually confirmed before execution, with a confirmation timeout set to 30 seconds. The historical instruction query function supports retrieving past maintenance records by time range. The performance evaluation of the decision tree uses a ten-fold cross-validation method, dividing the training set into ten parts that are used alternately as the test set. Evaluation metrics include classification accuracy, recall, and F1 score, and the evaluation results are used to guide the adjustment of the decision tree structure. Model optimization focuses on improving the ability to identify rare fault modes. Communication between the decision execution module and the upper-level monitoring system adopts the IEC 61850 protocol, which defines a comprehensive substation automation communication model. The data upload function periodically transmits maintenance instruction execution status and equipment status information, with a configurable transmission cycle. The remote debugging interface supports online modification and maintenance of decision tree parameters, with hierarchical management of modification permissions. The reliability design of the decision execution module adopts a dual-CPU hot standby redundancy scheme, automatically switching to the standby CPU in the event of a primary CPU failure. A watchdog timer monitors the program's running status, with a timeout set to 500 milliseconds. The power module supports dual DC power supplies, ensuring continuous system operation in the event of a single-channel failure.

[0085] Example 5: When acquiring surface infrared thermal imaging data in the status acquisition module, a dual-band infrared sensor is used to simultaneously acquire long-wave and mid-wave infrared images. The long-wave infrared detector of the dual-band infrared sensor has a response wavelength range of 8-14 micrometers, while the mid-wave infrared detector covers the 3-5 micrometer band. Synchronous acquisition is achieved by hardware triggering signals to expose both detectors simultaneously. The exposure time is dynamically adjusted between 100 microseconds and 10 milliseconds based on the target temperature. A non-uniformity correction algorithm is used to eliminate detector response differences. This algorithm employs a two-point correction method combined with scene-based correction, calibrated using a blackbody radiation source at both high and low temperature points. Pixel-level fusion is performed on the dual-band images using a weighted average method combined with Laplacian pyramid decomposition. The weighting coefficients are dynamically adjusted based on the signal-to-noise ratio. Composite thermal imaging data with a temperature sensitivity gradient is generated. This gradient is reflected in the difference in the rate of change of grayscale values ​​across different temperature ranges. Logarithmic mapping is used to enhance detail contrast in high-temperature regions. During the acquisition of mechanical displacement, a six-axis inertial measurement unit (IMU) and a laser displacement sensor are integrated. The IMU includes a three-axis accelerometer and a three-axis gyroscope, with an accelerometer range of ±16g and a gyroscope range of ±2000 degrees / second. The laser displacement sensor uses the triangulation principle, achieving a measurement accuracy of 0.01 mm and a sampling frequency of 1 kHz. The three-dimensional motion trajectory of the hard platen is recorded with millimeter-level accuracy. This trajectory is reconstructed using a sensor data fusion algorithm, which employs an extended Kalman filter to eliminate measurement noise.

[0086] The dual-band infrared sensor is installed 50 cm from the surface of the rigid pressure plate, with the installation angle perpendicular to the plate plane. The long-wave infrared image resolution is 640×512 pixels, and the mid-wave infrared image resolution is 320×256 pixels. The blackbody calibration temperature points for non-uniformity correction are set to 50℃ and 150℃, and the calibration process is automatically performed every 24 hours. Correction parameters for detector response differences are stored in a calibration file, which is updated in real time with temperature changes. The pixel-level fusion Laplacian pyramid decomposition has 5 layers, and the decomposition process uses Gaussian convolution kernels for downsampling. The weight calculation for the weighted average method is based on the standard deviation of a local region, defined as a 7×7 pixel window. The temperature sensitivity gradient is implemented through piecewise linear transformation, mapping the original temperature value to a 256-level grayscale range.

[0087] The six-axis inertial measurement unit (IMU) is mounted close to the center of the rigid pressure plate's rotating shaft, and the mounting base features a vibration-resistant design. The laser displacement sensor uses a 650 nm laser wavelength with a 0.5 mm spot diameter. Data synchronization employs hardware timestamps with an accuracy of 10 microseconds. The extended Kalman filter state variables for 3D motion trajectory reconstruction include position, velocity, acceleration, and Euler angles. The surface infrared thermal imaging data acquisition process includes three steps: detector preheating, autofocus, and image acquisition. The detector preheating time is no less than 30 minutes to ensure temperature stability. Autofocus is based on a contrast detection algorithm, with a focusing range from 20 cm to infinity. The image acquisition trigger signal is linked to the rigid pressure plate operating mechanism, acquiring 10 frames of images before and after the pressure plate's movement. The timing control of mechanical displacement acquisition is synchronized with the power grid frequency, with sampling points evenly distributed within each power frequency cycle. The six-axis IMU's data output frequency is 500 Hz, and the laser displacement sensor's sampling frequency is 1 kHz. Sensor data is transmitted via a CAN bus with a baud rate of 1 Mbps. The data buffer employs a circular buffer structure, storing 10 seconds of continuous data. The dual-band infrared sensor achieves a temperature measurement accuracy of ±1℃ and a spatial resolution of 0.5 milliradians. The residual noise after non-uniformity correction is less than 0.05℃. The image entropy value of pixel-level fusion is 15% higher than that of single-band fusion. The temperature sensitivity of composite thermal imaging data reaches 0.02℃ in the 20-100℃ range.

[0088] The six-axis inertial measurement unit (IMU) exhibits an acceleration measurement error of less than 0.1%FS and a gyroscope zero-bias stability of 0.5 degrees / hour. The laser displacement sensor has a linearity error of ±0.1%FS. The reconstruction accuracy of the 3D motion trajectory reaches 0.1 mm, and the angle measurement error is less than 0.1 degrees. The mechanical structure of the status acquisition module includes a sensor bracket, a heat dissipation system, and a protective housing. The sensor bracket is made of aluminum alloy and features a fine-tuning knob for precise alignment. The heat dissipation system includes heat sinks and a silent fan, ensuring normal operation of the sensor in ambient temperatures ranging from -40℃ to +85℃. The protective housing has an IP67 protection rating and its electromagnetic interference protection design meets the IEC61000 standard. The data acquisition software employs a multi-threaded architecture, including an acquisition thread, a processing thread, and a storage thread. The acquisition thread is responsible for sensor driving and data reading, the processing thread executes data preprocessing and fusion algorithms, and the storage thread writes data to the solid-state drive (SSD). The SSD has a storage capacity of 2TB and supports continuous storage of 30 days of data. The calibration and maintenance process includes daily automatic calibration and quarterly manual calibration. Daily automatic calibration performs non-uniformity correction and sensor zero-point calibration. Quarterly manual calibration verifies measurement accuracy using standard metrology equipment, and the calibration results generate a calibration report. The calibration report includes information on measurement error, uncertainty, and calibration date.

[0089] The status acquisition module features a wide voltage input design, with an input voltage range of DC12-36V. Power consumption is controlled through dynamic power adjustment, with standby power consumption less than 5W and full-power operation power consumption of 25W. Power protection includes overvoltage protection, undervoltage protection, and reverse connection protection. Data transmission interfaces include a Gigabit Ethernet port and a fiber optic interface. The Gigabit Ethernet port is used for local configuration and data export, while the fiber optic interface is used for long-distance transmission. Protocols supported include Modbus TCP and IEC61850. Modbus TCP is used for communication with monitoring systems, and IEC61850 is used for substation automation system integration. The status acquisition module's environmental adaptability meets the requirements of harsh industrial environments, with an operating temperature range of -40℃ to +70℃ and a relative humidity range of 5% to 95% without condensation. Vibration tolerance complies with IEC60068-2-6 standard, and shock tolerance complies with IEC60068-2-27 standard. Electromagnetic compatibility meets the IEC61000-6 series standards. Reliability design employs redundant power supplies and a watchdog timer. The redundant power supply supports hot-switching, and the watchdog timer has a 30-second timeout period. The self-diagnostic function monitors sensor status, communication links, and storage space; abnormal conditions trigger alarm signals. Alarm signals are output via dry contacts and simultaneously uploaded to the monitoring system.

[0090] The installation and commissioning of the status acquisition module includes mechanical installation, electrical wiring, and software configuration. Mechanical installation requires the sensor optical axis to be perpendicular to the plane of the hard platen, with an installation tolerance of less than 0.5 degrees. Electrical wiring uses shielded twisted-pair cable, with the shield grounded at a single point. Software configuration includes IP address settings, sampling parameter adjustments, and communication protocol selection. Performance testing includes static and dynamic tests. Static testing verifies measurement stability when the hard platen is stationary, while dynamic testing verifies trajectory tracking capability during operation of the hard platen. Test reports include measurement error, repeatability, and response time metrics. Maintenance includes regular cleaning of the optical window, connector inspection, and firmware updates. Optical window cleaning uses anhydrous ethanol and a lint-free cloth; connector inspection includes pin checks and insulation resistance measurements. Firmware updates support both remote online upgrades and local USB upgrades.

[0091] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0092] 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 relay protection hard-plate status intelligent monitoring system, characterized in that, include: The status acquisition module is used to acquire the mechanical displacement, contact resistance value and surface infrared thermal imaging data of the relay protection hard plate in real time. The feature parsing module is used to perform dynamic trajectory decomposition of mechanical displacement using a parallel sparse coding network to generate displacement feature vectors. At the same time, it extracts the gradient distribution pattern of contact resistance values ​​through a multi-scale residual convolutional network to generate a resistance feature matrix. Based on an adaptive threshold segmentation algorithm, it performs contour marking on high-temperature areas in infrared thermal imaging data to generate a thermodynamic feature map. The joint characterization module is used to input displacement feature vectors, resistance feature matrices and thermodynamic feature maps into a cascaded attention fusion network, correct feature dimension differences through a spatial alignment mechanism, and introduce learnable weight allocation coefficients in the channel dimension to output a multi-source joint state characterization. The anomaly prediction module is used to embed the multi-source joint state characterization into a lightweight time-series prediction model of contact physical constraints, and output the predicted values ​​of displacement offset, resistance attenuation rate and hot spot diffusion trend in the next three working cycles. The decision execution module is used to generate a set of hard plate maintenance instructions based on the predicted values ​​of displacement offset, resistance attenuation rate, and hot spot diffusion trend. The instruction set includes mechanical lubrication level, contact grinding frequency, and cooling fan speed adjustment.

2. The intelligent monitoring system for the status of relay protection hard plate according to claim 1, characterized in that, When the feature parsing module uses a parallel sparse coding network to perform dynamic trajectory decomposition of mechanical displacement: The time series of mechanical displacement is divided into overlapping sliding windows. The data in each window is learned by basis function through three sparse coding layers. The first layer extracts millisecond-level vibration components, the second layer separates second-level stroke fluctuation components, and the third layer fuses the outputs of the first two layers to generate a displacement feature vector. When processing contact resistance values ​​using a multi-scale residual convolutional network, five sets of convolutional kernels with different widths are used to scan the resistance value sequence in parallel. The output of each set of convolutional kernels is compressed by depth-separable convolution and then concatenated with the original gradient features passed by skip connections to form a resistance feature matrix.

3. The intelligent monitoring system for the status of relay protection hard-plate according to claim 2, characterized in that, When using the adaptive threshold segmentation algorithm to outline high-temperature regions in infrared thermal imaging data: First, bilateral filtering is used to eliminate thermal imaging noise. Then, the local entropy value of the image is calculated to dynamically adjust the segmentation threshold. For pixel clusters that exceed the threshold, morphological closing operation is performed to connect the broken edges. Finally, the closed polygon contour is extracted by the edge tracking algorithm and the centroid coordinates are marked to form a thermodynamic feature map.

4. The intelligent monitoring system for the status of relay protection hard plate according to claim 3, characterized in that, The operations of the cascaded attention fusion network in the joint representation module include: The displacement feature vector is spatially stacked with the resistance feature matrix after being transformed from one-dimensional to two-dimensional, and the resolution is adjusted by hollow spatial pyramid pooling. A region proposal network is used to generate regions of interest from the thermodynamic feature map and map them to the same spatial scale as the displacement-resistance hybrid feature. The contribution weights of the three types of features are dynamically calculated using gated cyclic units in the channel dimension, and the final output is a multi-source joint state representation that includes spatial alignment features.

5. The intelligent monitoring system for the status of relay protection hard plate according to claim 4, characterized in that, The lightweight time-series prediction model in the anomaly prediction module is constructed as follows: In the model encoder, grouped convolutions are used to reduce the number of parameters, and long short-term memory skip connections are introduced in the decoder. The physical constraints of the contact are transformed into regularization terms in the model loss function. The constraints include the inverse relationship between contact pressure and resistance, and the linear correlation between displacement and heat dissipation area. The model output layer is connected in parallel to three regression heads, which correspond to the predicted values ​​of displacement offset, resistivity attenuation rate, and hot spot diffusion trend, respectively.

6. The intelligent monitoring system for the status of relay protection hard plate according to claim 5, characterized in that, The process by which the decision execution module generates the hard plate maintenance instruction set includes: A decision tree containing twenty-four fault modes was established, with each branch node corresponding to different combinations of predicted values ​​for displacement offset, resistance attenuation rate, and hot spot diffusion trend. Preset maintenance parameters are stored in the leaf nodes. When the predicted value combination falls within the coverage area of ​​a certain leaf node, the corresponding mechanical lubrication level, contact polishing frequency and cooling fan speed adjustment command is triggered.

7. The intelligent monitoring system for the status of relay protection hard-plate according to claim 6, characterized in that, When acquiring surface infrared thermal imaging data in the status acquisition module: A dual-band infrared sensor is used to simultaneously acquire long-wave and mid-wave infrared images, and a non-uniformity correction algorithm is used to eliminate detector response differences. Pixel-level fusion of dual-band images generates composite thermal imaging data with temperature sensitivity gradients; During the mechanical displacement acquisition process, a six-axis inertial measurement unit and a laser displacement sensor are integrated to record the three-dimensional motion trajectory of the hard platen with millimeter-level accuracy.

8. The intelligent monitoring system for the status of relay protection hard plate according to claim 7, characterized in that, The structure of the multi-scale residual convolutional network in the feature parsing module includes: The first scale uses a convolution kernel with a width of 5 to capture the macroscopic trend changes in resistance values; The second scale uses a convolutional kernel with a width of 3 to extract local mutation features; The third scale expands the receptive field through dilated convolution with a dilation rate of 2; After being processed by the batch normalization layer, the outputs at each scale are fused element-wise with the feature maps passed through the cross-scale jump connections.

9. The intelligent monitoring system for the status of relay protection hard-plate according to claim 8, characterized in that, The spatial alignment mechanism in the joint characterization module is implemented as follows: The displacement eigenvectors are extended to a two-dimensional mesh using cubic spline interpolation. Fill the blank areas of the resistance feature matrix with Laplace diffusion results based on neighboring resistance values; By adjusting the geometric deformation of the thermodynamic feature map using deformable convolutional networks, the key regions are made consistent with the spatial distribution of displacement-resistance features.

10. The intelligent monitoring system for the status of relay protection hard-plate according to claim 9, characterized in that, The process of constructing the fault mode decision tree in the decision execution module includes: Five thousand sets of predicted values ​​and actual maintenance parameters were collected from historical maintenance records as training samples. The Gini coefficient is used as the splitting criterion for recursive feature partitioning. After eliminating overfitting branches using a pruning algorithm, each leaf node is bound to the standard operating procedure in the equipment maintenance manual.

Images