A transformer drying data acquisition processing system and method

Abstract

The application discloses a kind of acquisition and processing system and method of transformer drying data, it is related to data processing technical field;Including steps: step one, real-time acquisition and edge pre-processing of multi-source data;Step two, dry end point prediction model construction;Step three, composite objective function is constructed, and is solved using multi-objective evolutionary algorithm;Step four, real-time correction and feedback;The application monitors the key parameters in drying process in real time, automatically constructs accurate dry end point prediction model, and according to instant working condition and production strategy dynamic optimization heating and pumping curve, instead of traditional manual blind judgment and static preset;While significantly reducing operating energy consumption, through multi-objective evolutionary algorithm and online kalman filter correction closed-loop mechanism, the best balance of energy consumption and production efficiency is realized, not only improve drying quality and equipment safety, also greatly save energy cost, meet the dual needs of flexibility and energy saving and cost reduction of modern intelligent manufacturing.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a system and method for collecting and processing transformer drying data. Background Art

[0002] Drying has a decisive impact on the insulation performance, service life, and operational safety of transformers. It is a fundamental step in ensuring transformer quality. If drying is inadequate, residual moisture will reduce the insulation breakdown voltage and intensify partial discharge, shortening the transformer's service life and increasing the risk of accidents. Therefore, the drying process of transformers is particularly important.

[0003] Existing drying processes mainly rely on fixed parameters, and the drying endpoint needs to be manually judged. This method is rigid and lacks flexibility, and does not meet the needs of modern production technology. At the same time, traditional drying processes rely on long-term high-temperature heating and vacuum extraction. This static preset mode causes a lot of energy consumption and increases costs. Summary of the Invention

[0004] The main purpose of the present invention is to provide a system and method for collecting and processing transformer drying data to overcome the problems mentioned in the above background technology.

[0005] To achieve the above objectives, according to one aspect of the present invention, a system for collecting and processing transformer drying data is provided, the system comprising:

[0006] The multi-source acquisition and processing module communicates with all sensors in the drying process to collect five-dimensional dynamic data, including temperature, vacuum level, dew point temperature, vibration intensity, and instantaneous power. In addition, the multi-source acquisition and processing module also stores static process data of the drying process, including coil resistance, insulation layer thickness, magnetic permeability, gap width between laminations, and ambient temperature and humidity. The multi-source acquisition and processing module is also used to pre-process the raw data, specifically:

[0007] Timing synchronization: Using whole seconds as the time unit, first downsample the high-frequency signals of temperature and vibration to once per second; then perform linear interpolation on the low-frequency signal of dew point to align all signals to the same time point;

[0008] Outlier point removal: Calculate the historical mean and standard deviation of each time series. If the reading at a certain moment deviates from the mean by more than three times the standard deviation, the point is marked as an anomaly and removed to prevent instantaneous interference from affecting subsequent analysis.

[0009] Real-time feature construction: At each timestamp, the first-order difference of physical quantities and the gradient change of dew point in space are calculated. This allows the above original parameters to be time-series aligned and abnormal data to be eliminated to obtain clearly organized data, including temperature, temperature change rate, vacuum degree, vacuum change rate, dew point, dew point gradient, vibration intensity, and instantaneous power.

[0010] Data standardization: To eliminate differences in dry tank volume and insulation material specific heat capacity across different factories or workshops, each time series feature is first scaled by the inverse of the tank's actual volume, and then rescaled by the material's thermal conductivity. This allows all features to be mapped to the same standard volume and specific heat capacity, ensuring consistent and directly comparable model input dimensions.

[0011] The prediction model building module includes an attention recognition unit, a temporal feature capture unit, and a fully connected unit, which is used to build a drying endpoint prediction model based on preprocessed data. Specifically:

[0012] The attention recognition unit uses multi-head attention to automatically filter out the most critical time nodes and fluctuation patterns for determining the drying endpoint from the five-dimensional time series input and outputs a sequence code;

[0013] Capturing time series feature units, using multi-layer long short-term memory networks based on sequence encoding to capture long-term dependencies in time series features to control information retention and updating;

[0014] The fully connected unit extracts a vector representing the evolution of the entire process from the last hidden state of the top-level LSTM and concatenates it with the vector s of static process parameters, including coil resistance, insulation thickness, magnetic permeability, lamination gap, and average ambient humidity, in the feature dimension to form a fused vector z. After passing through a multi-layer fully connected network and activation, it ultimately outputs three regression indicators: remaining drying time, estimated energy consumption, and drying quality score.

[0015] A multi-objective solution module is used to minimize the total drying time and minimize the cumulative energy consumption as the dual objectives and embed the physical constraints of dew point, vacuum change rate and temperature upper limit, and use a multi-objective evolutionary algorithm to generate a Pareto frontier solution;

[0016] The real-time correction feedback module is used to fuse the model prediction and the actual completion time residual through a one-dimensional Kalman filter, and trigger the local fine-tuning mechanism when the residual exceeds the threshold to dynamically correct the control curve and close the loop.

[0017] Further, the sequence encoding output process:

[0018] Let the time series input matrix be , where T is the total number of time steps, d is the feature dimension of each step, and the query matrix is ​​generated by projecting it into the subspace of query, key and value through three sets of trainable linear mappings , key matrix Sum Matrix , where each row corresponds to a time step, 、 、 ∈ , is the trainable data, mapping maintenance is often taken ; For the query component at any time i with the key vector at all moments j , calculate the similarity score by dot product and dividing by the scaling factor : , measures the similarity between the state of the current moment i and the historical moment j, for example, whether the temperature, vacuum degree and other characteristics of the two moments show the same change pattern; then, all scores are normalized to attention weights using the softmax function under the same query i , is the attention weight of the information at moment i to moment j, Indicates how much information should be “borrowed” from the past seconds when predicting the remaining drying time and energy consumption at the i-th second. The larger the attention weight, the more significant the impact of the corresponding historical moment on the current decision; j' is only a dummy index, which is used to distinguish j in the numerator from the traversal index in the summation process and does not represent a new physical quantity; for each moment i, all weighted values ​​corresponding to the attention weight are used. And sum it up to get the attention output : ,because The sum of the values ​​in the same row is 1, and the j with the highest weight is the most representative key moment or fluctuation node. Represents the attention output vector of the i-th second, which is the vector of all values In order to capture multiple attention patterns, h groups of parallel heads are actually used, each learning an independent 、 、 Finally, the outputs of h heads are concatenated in the feature dimension and linearly mapped again to form an attention encoding sequence of uniform dimension. ,in ;

[0019] In order to capture multiple attention patterns, h sets of parallel heads are actually used, each learning an independent query weight matrix, key weight matrix and value weight matrix. Finally, the outputs of the h heads are spliced ​​in the feature dimension and linearly mapped again to become an attention encoding sequence of uniform dimension. ,in .

[0020] Furthermore, the process of capturing long-term dependencies in time series features is as follows:

[0021] In obtaining the attention encoding sequence After that, it is input into a multi-layer long short-term memory network (LSTM) to further extract the dynamic evolution features across time windows. For each time step t, the LSTM maintains a cell state and hidden state and through the forget gate , input gate AND output gate To control the retention and updating of information: , where σ is Sigmoid activation, is element-by-element multiplication, Wf, Wi, Wc, Wo are trainable weights, due to the cell state It can be transmitted across multiple time steps without attenuation. Important early information can continue to influence the network output later in the sequence through the forget gate and input gate, thereby capturing slow trends in the drying process, periodic fluctuations, or long-term dependencies such as equipment aging. When deeper temporal abstraction is required, multiple LSTM layers can be stacked, with the upper layer input being the hidden state [h1, h2, …, hT] of the lower layer, to obtain richer multi-scale temporal representations.

[0022] Furthermore, the process of the fully connected network outputting three regression results is:

[0023] After the time series modeling is completed, the hidden vector hT at the last moment is extracted from the hidden state sequence of the top-level LSTM. This vector aggregates the dynamic evolution information from t=1 to T. At the same time, the system also collects a set of static process parameters related to the hardware structure and environmental conditions—including coil resistance, insulation layer thickness, magnetic permeability, lamination gap, and average ambient humidity—to form a static feature vector s. By splicing hT and s in the feature dimension: , and the final fusion vector z is obtained. The fusion vector then enters the multi-layer fully connected network, and each layer is in the form of ,in ReLU activation is usually selected, and the last layer outputs a vector y=[y1,y2,y3] of length 3, where y1, y2, and y3 correspond to the remaining drying time, expected energy consumption, and drying quality score, respectively. To ensure physical rationality, the model imposes boundary constraints on the output: y1 ≥ 0, y2 ≥ 0, and y3∈[0,1]. In this way, dynamic timing information is deeply integrated with the properties of the equipment itself to achieve accurate multi-objective regression prediction of the drying process.

[0024] Furthermore, the specific process of the multi-objective evolutionary algorithm is as follows:

[0025] 3-1, extract the temperature curve D(t) and vacuum curve P(t). The two curves are usually expressed in the form of discrete control nodes. The composite objective function aims to minimize the total drying time and total energy consumption at the same time. The composite objective function is constructed and its expression is: ,in The total time required to complete one drying cycle is The accumulated power consumption to complete one drying process;

[0026] The constraints of the composite multi-objective function are: 1. Dew point constraint: at the drying end point, the dew point temperature does not exceed -40°C; 2. Vacuum change rate constraint: the vacuum change rate does not exceed 20 Pa per minute; 3. Temperature upper limit constraint: the temperature does not exceed 110°C;

[0027] 3-2, randomly generate several groups of candidate solutions, each group consists of the same number of control nodes, to ensure that all candidate initial solutions meet the constraints; each individual in the population explores the optimal balance through different curve combinations; for each candidate solution, call the fast simulation of the drying endpoint prediction model constructed above to predict when the dew point will reach -40℃ for the first time under the control of the temperature curve, and predict the total drying time required in this cycle as , and predict the cumulative energy consumption during the period If the temperature or vacuum rate is not constrained during the simulation, a penalty is imposed on the candidate solution, thereby obtaining the dual objective value of each individual , where n is a positive integer, representing the index of the candidate solution;

[0028] 3-3, find the optimal control strategy based on the composite objective function and constraints of step 3-1.

[0029] Furthermore, the process of finding the optimal control strategy is:

[0030] The population is stratified, and the first layer is the Pareto frontier: any solution in this layer has a and total energy consumption Neither of the two objectives can be dominated by other solutions in the same generation at the same time; the second layer is composed of solutions that are dominated by the first layer solutions but do not have a complete domination relationship with each other within the second layer; and so on, the entire population is divided into multiple non-dominated layers;

[0031] Within each layer, the crowding degree of each solution in the two-objective space is calculated, that is, the distance between the solution and its neighboring solutions on the objective plane. The specific strategy is: 1. Prioritize the retention of non-dominated layers closer to the front; 2. Within the same layer, prioritize the retention of individuals with high crowding degrees. Next, a pairwise crossover is used from the retained pool to exchange some of the control nodes of the parent solution to generate a new generation of candidate solutions, thereby inheriting and mixing the outstanding time period control strategies of the parent.

[0032] The above process is iterated continuously, and the drying time, energy consumption and quality penalty are recalculated in each generation. The new solutions are sorted by non-domination and screened by congestion until the Pareto front is stable or the preset maximum number of iterations is reached. When the algorithm converges, a set of mutually non-dominated optimal solutions is obtained, and each solution corresponds to a complete temperature-vacuum control curve and its drying time and energy consumption. Balanced results; according to the control scheme selected by the user or production scheduler, the corresponding node set values ​​are sent to the heater and vacuum pumping system through the interface in real time to form a closed-loop automatic adjustment.

[0033] In order to achieve the above object, according to another aspect of the present invention, a method for collecting and processing transformer drying data is provided, the method comprising the following steps:

[0034] Step 1: Communicate with all sensors in the drying process to collect five-dimensional dynamic data, including temperature, vacuum level, dew point temperature, vibration intensity, and instantaneous power. In addition, the multi-source acquisition and processing module also stores static process data of the drying process, including coil resistance, insulation layer thickness, magnetic permeability, gap width between laminations, and ambient temperature and humidity.

[0035] Step 2: Extract time series dynamic data and static process data to build key prediction models;

[0036] Step 3: Based on the dual objectives of minimizing the total drying time and minimizing the cumulative energy consumption, a composite objective function that takes into account both time efficiency and energy consumption minimization is established, and a multi-objective evolutionary algorithm is used to solve it;

[0037] In step 4, the model prediction and the actual completion time residual are fused through a one-dimensional Kalman filter, and when the residual exceeds the threshold, the local fine-tuning mechanism is triggered to dynamically correct the control curve and close the loop.

[0038] Beneficial effects of the present invention:

[0039] By inputting multi-channel time-series dynamic data and key static process parameters into attention, LSTM, and fully connected deep network structures, the present invention enables the model to automatically focus on key moments or fluctuation nodes that affect endpoint judgment, while being compatible with the inherent differences in the physical properties of the equipment. Multi-head attention provides the ability to capture fluctuations across angles, LSTM mines slow-changing trends across long time ranges, and the parallel fusion of static features enriches the model's perception of material and structural differences. The prediction model constructed in this way is both accurate and reliable in estimating remaining drying time, expected energy consumption, and drying quality, and has good robustness. It can continuously provide high-fidelity endpoint prediction and energy consumption assessment in complex and changing industrial environments.

[0040] This invention uses a multi-objective evolutionary algorithm to optimize the process by minimizing total drying time and cumulative energy consumption as the dual objectives, combined with physical safety constraints such as dew point, vacuum rate, and temperature upper limit. This ensures a dynamic balance between high production capacity and low energy consumption while fully maintaining process and equipment safety. Simultaneously, a series of mutually independent optimal solutions are generated on the Pareto frontier. The system can instantly adapt to various production needs and distribute the preferred solution to the site, achieving intelligent closed-loop scheduling and real-time visual decision support.

[0041] The present invention uses Kalman filtering to perform weighted fusion of the model prediction and the endpoint time residual of field observation, effectively solving the nonlinear disturbance and measurement noise of the industrial system, making the state estimation both sensitive and responsive and smooth and stable; when the residual exceeds the set threshold, the system will perform local small-step fast adjustment during the period when the error is most significant, slightly increase the temperature node or increase the pumping intensity, and strictly limit it to the upper limit of the insulation material temperature and the vacuum change rate, which can not only quickly reduce the prediction deviation, but also avoid severe impact on the overall control, ensuring the long-term optimization, safety and reliability of the drying process, and realizing the adaptive correction of the model to the field working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of system module connection of the present invention;

[0043] Figure 2 It is a schematic flow chart of the method of the present invention. DETAILED DESCRIPTION

[0044] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0045] like Figure 1As shown, this embodiment provides a transformer drying data collection and processing system, which includes: a multi-source collection and processing module, a prediction model building module, a multi-objective solution module and a real-time correction feedback module;

[0046] The multi-source acquisition and processing module collects temperature, vacuum level, dew point temperature, vibration intensity, and instantaneous power in real time by deploying distributed fiber-optic temperature sensors in key locations such as the transformer core and coils, installing piezoresistive vacuum gauges and chilled mirror dew point meters in the drying chamber, and deploying high-sensitivity vibration sensors and smart meters on the vacuum pump body. These five types of heterogeneous data are first accurately timestamped to the millisecond level as they flow through the edge gateway. The time series is then unified using a synchronization method of downsampling to the entire second and linear interpolation. The three-sigma principle is then used to eliminate transient mutation points, effectively filtering out mechanical jitter and environmental interference. The gateway then calculates first-order features such as the instantaneous rate of change of temperature and vacuum level, and the spatial gradient of the dew point, at every entire second. Finally, to eliminate the heterogeneity caused by tank volumes and insulation material heat capacities across different plant locations, the system normalizes all time series features using a dual factor: the inverse of the tank volume and the material heat capacity coefficient. This ensures that model inputs are consistently dimensional and directly comparable, regardless of site dimensions and material variations.

[0047] The warning model building module includes an attention recognition unit, a unit for capturing temporal features, and a fully connected unit, an attention recognition unit;

[0048] The attention recognition unit uses multi-head attention to automatically filter out the most critical time nodes and fluctuation patterns for determining the drying endpoint from the five-dimensional time series input. Specifically, a trainable linear mapping is used to project the features of each time step into a query, key, and value vector. The dot product similarity between the query and all historical keys is calculated, and the attention weight is then obtained through scaling and softmax normalization. This determines the historical moment that provides the most valuable information for predicting the remaining time and energy consumption in the i-th second. Multiple attention heads learn association patterns from different angles in parallel, and their outputs are then spliced ​​and linearly mapped back to the original dimension, forming a highly compressed sequence encoding that focuses on high-value moments.

[0049] After obtaining the attention coding sequence, the temporal feature capture unit is input into a multi-layer long short-term memory network to capture slow trends and periodic fluctuations spanning tens of minutes or even hours. Through the gating mechanism of forget gates, input gates, and output gates, early key features are selectively retained and irrelevant information is suppressed, allowing the model to continuously utilize important early states in the later stages of the sequence. This structure is particularly adept at handling nonlinear long-term dependencies caused by material aging, equipment performance changes, or environmental fluctuations during the drying process. By stacking multiple layers of LSTM, the model can refine dynamic evolution features at different time scales, further enhancing the ability to express complex temporal patterns and providing a robust temporal basis for drying endpoint and energy consumption prediction.

[0050] After temporal encoding is complete, the fully connected unit extracts a vector representing the evolution of the entire process from the last hidden state of the top-level LSTM. This vector is concatenated with the vector s containing static process parameters such as coil resistance, insulation thickness, magnetic permeability, lamination gap, and average ambient humidity in the feature dimension to form a fused vector z. After passing through a multi-layer fully connected network and ReLU activation, this vector ultimately outputs three regression indicators: remaining drying time, estimated energy consumption, and drying quality score.

[0051] After predicting the endpoint timing of the drying process, the multi-objective solution module sets the dual objectives of minimizing total drying time and cumulative energy consumption. A composite objective function is established by performing evolutionary optimization on discrete nodes of the heating temperature and vacuum curves. To ensure process safety, the optimization process incorporates three physical constraints: the endpoint dew point must be below -40°C, the vacuum rate of change must not exceed 20 Pa / min, and the temperature limit is 110°C. Initially, a population of candidate solutions that meet these constraints is randomly generated. Subsequently, in each generation, the drying endpoint prediction model is used to rapidly estimate the drying time and energy consumption of each solution, with penalties applied to solutions that violate these constraints. A non-dominated sorting and congestion-preserving strategy is used to select the Pareto front solution set. Crossover and micro-mutation operations are then used to blend excellent fragments between parent generations and expand the solution space. Once the algorithm iterates until the Pareto front stabilizes, users can freely select the optimal control curve based on their current production priority (capacity or energy conservation) from a two-dimensional visualization of drying time and energy consumption. This curve is then distributed to the site for real-time execution, achieving closed-loop optimization scheduling that prioritizes both energy efficiency and equipment safety.

[0052] After each prediction is completed, the real-time correction feedback module compares the model's predicted drying endpoint time with the actual completion time when the on-site dew point remains below -40°C and vacuum and power changes are stable. The module calculates the residual to quantify the deviation between the prediction and the actual operating conditions. Taking into account the nonlinearity and measurement noise of the drying process, the system introduces a one-dimensional Kalman filter. Without fully trusting the prediction or relying entirely on observation, the system weights and fuses the two based on the degree of trust (error covariance), thereby dynamically correcting the state estimate so that the prediction results are both responsive to on-site changes and smooth and stable. When the residual exceeds the set threshold, a local fine-tuning mechanism is triggered to accurately compensate for the model error. This local fine-tuning mechanism can quickly converge the prediction and target with minimal disturbance within the critical period without restarting the global optimization, ensuring that the drying control remains continuously optimal, safe, and reliable.

[0053] According to an embodiment of the present invention, Figure 2 As shown, a method for collecting and processing transformer drying data is also provided, which includes the following steps:

[0054] Step 1: Real-time collection and edge preprocessing of multi-source data;

[0055] Distributed fiber optic temperature sensors, tens of meters long, are evenly laid across key locations, such as the transformer core and coils. These sensors continuously sense temperature distribution along the length of the fiber, are highly resistant to electromagnetic interference, and collect temperature data in real time at a rate of 10 times per second. A piezoresistive vacuum gauge is installed within the vacuum drying chamber, covering a range from atmospheric pressure (101 kPa) to deep vacuum, providing real-time information on the vacuuming process. This sensor collects pressure readings once per second to assess vacuum level trends. A chilled mirror dew point meter measures the dew point of the air within the chamber. This instrument determines the dew point by cooling the mirror and detecting the temperature at which condensed water droplets first appear, covering a range from -70°C to +20°C. It outputs a dew point reading every five seconds. A highly sensitive vibration sensor is installed on the vacuum pump body, capturing vibration fluctuations caused by bearing anomalies and insufficient lubrication, monitoring the equipment's operating status at a rate of 10 times per second. A smart meter is connected to the power inlet to record the instantaneous power consumption of the heating and exhaust systems in real time, updating every second to provide a basis for energy consumption analysis.

[0056] The edge gateway continuously receives raw data streams from various sensors. It first affixes a millisecond-accurate timestamp to each time series signal. It then fuses the multi-source data through the following steps:

[0057] (1) Timing synchronization: Using the whole second as the time unit, first downsample high-frequency signals such as temperature and vibration to once per second; then perform linear interpolation on low-frequency signals such as dew point to align all signals to the same time point;

[0058] (2) Abnormal point elimination: Calculate the historical mean and standard deviation of each time series. If the reading at a certain moment deviates from the mean by more than three times the standard deviation, the point is marked as abnormal and eliminated to avoid the impact of instantaneous interference on subsequent analysis;

[0059] (3) Real-time feature construction: At each timestamp, the first-order differences of key physical quantities (such as the rate of change of temperature and vacuum) and the spatial gradient of dew point are calculated to enhance sensitivity to sudden fluctuations and local weak links; thus, the above original parameters can be aligned in time series, and abnormal data can be eliminated to obtain clearly organized data, including temperature (degrees Celsius), temperature change rate (degrees Celsius per second), vacuum (kilopascals), vacuum change rate (kilopascals per second), dew point (degrees Celsius), dew point gradient (degrees Celsius per meter), vibration (meters per square second), and instantaneous power (kilowatts);

[0060] (4) Data standardization: The drying tank volumes and insulation material heat capacities used in different factories or workshops vary, and direct use of the model will result in deviations. Therefore, it is necessary to map the data features to a uniform standard scale. First, the data is scaled according to the inverse of the actual volume of the tank, so that all data are equivalent to the standard volume. Then, a secondary adjustment is made according to the thermal melting coefficient of the material, so that the value corresponding to the standard heat capacity is the same. In this way, regardless of the differences in the tank size and material properties of the factory, the model input remains in the same dimension and can be directly compared and transferred.

[0061] By laying distributed fiber optic temperature sensors on the transformer core and coils, and combining them with high-frequency data acquisition using piezoresistive vacuum gauges, chilled mirror dew point meters, vibration sensors, and smart meters, the temperature distribution, vacuum degree, dew point, vibration conditions, and power consumption during the drying process can be accurately recorded with millisecond-level timestamps. Subsequently, timing synchronization, three-sigma anomaly removal, and real-time feature construction are performed on the edge gateway, and the tank volume and material heat capacity of different factories are standardized and compensated, which can significantly improve the integrity, consistency, and comparability of the data. This not only provides high-quality, noise-free, and dimensionally unified input for subsequent modeling, but also significantly reduces the risk of false alarms and missed alarms by eliminating sensor jitter and environmental interference in real time, ensuring that the system can stably and accurately reflect the drying status and equipment health under various operating conditions.

[0062] Step 2: Construction of drying endpoint prediction model;

[0063] Extract time-series dynamic data and static process data. Specific time-series dynamic parameters include: temperature, vacuum, dew point, vibration intensity, instantaneous power, and energy consumption with timestamps. Temperature curves, vacuum curves, dew point curves, vibration curves, instantaneous power curves, and energy consumption curves are constructed according to the corresponding timestamps. Static process data includes coil resistance, insulation layer thickness, core material magnetic permeability, gap width between laminations, and ambient temperature and humidity. These static process parameters do not change significantly over time, but they directly affect drying efficiency.

[0064] 2-1. The model first uses the attention mechanism to automatically identify the most representative moments and fluctuation patterns in the time series data; specifically:

[0065] We first let the time series input matrix be , where T is the total number of time steps, d is the feature dimension of each step (five-dimensional feature data of temperature, vacuum degree, dew point, vibration, and energy consumption), and the query matrix is ​​generated by projecting it into the subspace of query, key, and value through three sets of trainable linear mappings. , key matrix Sum Matrix , where each row corresponds to a time step, 、 、 ∈ , is the trainable data, mapping maintenance is often taken ; For the query component at any time i with the key vector at all moments j , calculate the similarity score by dot product and dividing by the scaling factor : , measures the similarity between the state of the current moment i and the historical moment j, for example, whether the temperature, vacuum degree and other characteristics of the two moments show the same change pattern; then, all scores are normalized to attention weights using the softmax function under the same query i , is the attention weight of the information at moment i to moment j, Indicates how much information should be “borrowed” from the past seconds when predicting the remaining drying time and energy consumption at the i-th second. The larger the attention weight, the more significant the impact of the corresponding historical moment on the current decision; j' is only a dummy index, which is used to distinguish j in the numerator from the traversal index in the summation process and does not represent a new physical quantity; for each moment i, all weighted values ​​corresponding to the attention weight are used. And sum it up to get the attention output : ,because The sum of the values ​​in the same row is 1, and the j with the highest weight is the most representative key moment or fluctuation node. Represents the attention output vector of the i-th second, which is the vector of all values In order to capture multiple attention patterns, h groups of parallel heads are actually used, each learning an independent 、 、 Finally, the outputs of h heads are concatenated in the feature dimension and linearly mapped again to form an attention encoding sequence of uniform dimension. ,in , which allows the model to extract temporal dependencies from multiple perspectives (such as temperature-dominant, vacuum-dominant, energy-dominant, etc.) in parallel, enhancing the ability to express complex drying processes;

[0066] 2-2, then the long-term dependencies in the time series features are captured through the long short-term memory network (LSTM), specifically:

[0067] In obtaining the attention encoding sequence After that, it is input into a multi-layer long short-term memory network (LSTM) to further extract the dynamic evolution features across time windows. For each time step t, the LSTM maintains a cell state and hidden state and through the forget gate , input gate AND output gate To control the retention and updating of information: , where σ is Sigmoid activation, For element-by-element multiplication, W f ,W i ,W c ,W o is a trainable weight, since the cell state It can be transmitted without attenuation across multiple time steps. Important early information can continue to affect the network output in the later stages of the sequence through the forget gate and input gate, thereby capturing the slow trend of the drying process, periodic fluctuations or long-term dependencies such as equipment aging. When a deeper level of temporal abstraction is required, multiple LSTM layers can be stacked, with the upper layer input being the hidden state of the lower layer [h1,h2,…,h T ], in order to obtain richer multi-scale temporal expressions;

[0068] 2-3. Finally, the obtained time series code is spliced ​​with the static process parameters and fed into the fully connected network to output three regression results: remaining drying time, estimated energy consumption, and overall drying quality score; specifically:

[0069] After the time series modeling is completed, the hidden vector h of the last moment is extracted from the hidden state sequence of the top LSTM T , which aggregates the dynamic evolution information from t=1 to T. At the same time, the system also collects a set of static process parameters related to the hardware structure and environmental conditions - including coil resistance, insulation layer thickness, magnetic permeability, lamination gap, and average ambient humidity - to form a static feature vector s. By splicing hT and s in the feature dimension: , and the final fusion vector z is obtained. The fusion vector then enters the multi-layer fully connected network, and each layer is in the form of ,in ReLU activation is typically chosen, and the final layer outputs a vector of length 3, y=[y1,y2,y3], where y1, y2, and y3 correspond to the remaining drying time, estimated energy consumption, and drying quality score, respectively. To ensure physical rationality, the model imposes bounds on the output: y1 ≥ 0, y2 ≥ 0, and y3 ∈ [0, 1]. This approach deeply integrates dynamic timing information with the device's inherent properties, enabling accurate multi-objective regression prediction of the drying process. It should be noted that during training, mean absolute error, root mean square error, and coefficient of determination are used as evaluation indicators, with 80% of the data used for training and 20% for validation.

[0070] By inputting multi-channel time-series dynamic data and key static process parameters into attention, LSTM and fully connected deep network structures respectively, the model can automatically focus on key moments or fluctuation nodes that affect endpoint judgment, while being compatible with the inherent differences in the physical properties of the equipment; multi-head attention provides the ability to capture fluctuations across angles, LSTM mines slow-changing trends over long time ranges, and the parallel fusion of static features enriches the model's perception of material and structural differences. The prediction model constructed in this way is both accurate and reliable in estimating the remaining drying time, expected energy consumption and drying quality, and has good robustness. It can continuously provide high-fidelity endpoint prediction and energy consumption assessment in complex and changing industrial environments.

[0071] Step three: Establish a composite objective that takes into account both time efficiency and energy consumption minimization, and use a multi-objective evolutionary algorithm to solve it. This not only ensures a balance between drying efficiency and energy saving, but also has good real-time adaptability and visual decision support; specifically:

[0072] 3-1. Extract the temperature curve and vacuum curve. During the entire drying cycle, the heating temperature at each time point t is represented by D(t). Technicians control the heat input rate in the cavity by adjusting the heating temperature D(t), which affects the water evaporation rate and drying time. The cavity pressure at each time point t is represented by P(t). The vacuum reflects the volatilization conditions of water in the gas phase. A lower P(t) helps to reduce the saturation pressure of water vapor and accelerate evaporation. It should be noted that these two curves are usually represented in the form of discrete control nodes. For example, a temperature and a vacuum set value are set every 5 minutes, and the curves have linear or smooth differences between the nodes. The composite objective function aims to minimize the total drying time and total energy consumption at the same time. Its expression is: ,in The total time required to complete one drying cycle is To complete the cumulative power consumption during a drying process; the constraints of this composite multi-objective function are: (1) dew point constraint, at the end of drying, the dew point temperature does not exceed -40 ° C; (2) vacuum change rate constraint, the vacuum change rate does not exceed 20 Pa per minute to avoid sudden changes and damage to the pump body; (3) temperature upper limit constraint, the temperature does not exceed 110 ° C to prevent overheating of the insulation material; physical and safety constraints are imposed on the solution space to ensure that the process is feasible and the equipment is protected;

[0073] 3-2, randomly generate several groups of candidate solutions , each group consists of the same number of control nodes to ensure that all candidate initial solutions meet the constraints; it represents the temperature and vacuum control curve corresponding to the nth candidate solution in the evolutionary algorithm, and each individual in the population explores the optimal balance through different curve combinations; for each candidate solution, the fast simulation of the drying endpoint prediction model constructed above is called to predict when the dew point can reach -40℃ for the first time under the control of the temperature curve, as , and predict the cumulative energy consumption during the period If the temperature or vacuum rate is not constrained during the simulation, a penalty is imposed on the candidate solution, thereby obtaining the dual objective value of each individual ;

[0074] 3-3, the population is stratified, the first layer is the Pareto frontier: any solution in this layer has a and total energy consumption Neither of the two objectives can be dominated by other solutions in the same generation at the same time; the second layer is composed of solutions that are dominated by the first layer solutions but do not have a complete domination relationship with each other within the second layer; and so on, the entire population is divided into multiple non-dominated layers; within each layer, the crowding degree of each solution in the two-objective space is calculated - that is, the distance between the solution and the adjacent solution on the target plane. The farther the distance (sparse area), the greater the crowding degree. This ensures that when selecting, not only the optimal performance solution is retained, but also diversity is taken into account; the specific strategy is: 1. Prioritize the retention of non-dominated layers at the front because they represent a better time-energy balance; 2. Within the same layer, prioritize the retention of individuals with higher crowding to avoid the concentration of solutions in local areas; next, use pairwise crossover from the retention pool to remove part of the parent solution. The sub-control nodes (temperature nodes and vacuum nodes) are exchanged with each other to generate a new generation of candidate solutions, thereby inheriting and mixing the period control strategies that perform well in the parent generation. In order to prevent the population from falling into local optimality too early, the temperature or vacuum curve nodes of the new solution are slightly randomly perturbed (mutated), and a small random offset is applied to the control values ​​of adjacent nodes to expand the search space; the above process is iterated continuously, and the drying time, energy consumption and quality penalty are recalculated in each generation, and the new solutions are non-dominated sorted and screened for congestion until the Pareto front is stable (the front solution set for several consecutive generations hardly changes) or the preset maximum number of iterations is reached; when the algorithm converges, a set of optimal solutions that do not dominate each other is obtained, and each solution corresponds to a complete temperature-vacuum control curve and its drying time and energy consumption. Balanced results; on a two-dimensional target plane (total drying time on the horizontal axis, total energy consumption on the vertical axis), these solutions are arranged into an approximate Pareto curve. Users or production schedulers can freely select the most appropriate control scheme on the leading-edge curve based on current production priorities (for example, prioritizing production capacity or energy conservation). Based on the selected optimal scheme, the corresponding node setpoints are sent to the heater and vacuum extraction system in real time through the interface, forming a closed-loop automatic adjustment.

[0075] By taking minimizing the total drying time and minimizing the cumulative energy consumption as the dual objectives, combined with physical safety constraints such as dew point, vacuum rate and temperature upper limit, and using a multi-objective evolutionary algorithm for optimization, we can ensure the dynamic balance between high production capacity and low energy consumption, and fully maintain the safety of process and equipment; at the same time, a series of mutually non-dominant optimal solutions are generated on the Pareto frontier. The system can achieve instant adaptation to various production needs (such as giving priority to energy saving during peak power consumption periods and giving priority to speed increase during emergency order periods), and send the preferred solution to the site to realize intelligent closed-loop scheduling and real-time visual decision support.

[0076] Step 4: Real-time Correction and Feedback: After each drying endpoint prediction is completed, the predicted drying endpoint time is compared with the actual observed endpoint in real time, and the time residual between the two is calculated. The endpoint is determined based on the following criteria: the dew point temperature is continuously and stably below -40°C, the vacuum level remains stable, and the energy consumption curve changes slowly. The physical significance of this residual is to quantitatively assess the error between the model prediction value and the actual system state, reflecting the degree of deviation of the prediction model under the current operating conditions. Due to the highly nonlinear and complex perturbation characteristics of industrial systems, it is difficult for any single model to maintain absolute accuracy in all operating stages. To gradually correct the error without interrupting the drying control process, a one-dimensional Kalman filter is introduced to dynamically fuse and correct the prediction results. The basic idea of ​​the Kalman filter is to obtain the optimal estimate of the current system state by taking a weighted average based on the trust level (i.e., the error covariance) of both the model prediction results and the observed data, without complete trust in either the model prediction results or the observed data. Specifically, the filter combines the model prediction values ​​with the actual observed data at each time point and continuously updates the state estimate through an iterative method, so that the prediction results gradually approach the actual operating conditions without excessive fluctuations, thereby improving the prediction stability and anti-interference ability.

[0077] When the residual exceeds the set threshold (for example, 5 minutes), the system will trigger the local fine-tuning mechanism, which performs local small-step quick adjustment based on the current control trajectory to correct key control parameters. First, the time period with the most significant positioning error, usually the time interval with the largest deviation between the predicted value and the observed value, is selected as the reference for the fine-tuning section. Then, at the first 2 to 3 temperature control nodes of the error section, the target temperature setting value is appropriately increased to increase the heat input per unit time, thereby accelerating the water evaporation rate and compressing the actual drying time. The fine-tuning process strictly meets two physical constraints: 1. The temperature must not exceed The upper temperature limit that the insulation material can withstand (e.g., 110°C). 2. The temperature rise rate between any two adjacent temperature control nodes must be less than the set maximum temperature growth rate (e.g., no more than 2°C per minute) to prevent local overheating and insulation damage. If the current temperature is close to the set upper limit, but the dew point continues to drop slowly, it indicates that the heat input has reached saturation, and the vacuum control parameters will be adjusted. The specific strategy is to appropriately reduce the target vacuum value (i.e., increase the pumping intensity), but the vacuum change rate must not exceed 20 Pa / minute to prevent the pumping system from being subjected to excessive shock loads and damaging equipment stability.

[0078] By using Kalman filtering to perform weighted fusion of the model prediction and the endpoint time residual of field observation, the nonlinear disturbance and measurement noise of the industrial system are effectively resolved, making the state estimation both sensitive and smooth and stable. When the residual exceeds the set threshold, the system will perform local small-step fast adjustment during the period when the error is most significant, slightly increasing the temperature node or increasing the pumping intensity, and strictly limited to the upper temperature limit of the insulating material and the vacuum change rate. This can not only quickly reduce the prediction deviation, but also avoid severe impact on the overall control, ensuring the long-term optimization, safety and reliability of the drying process, and realizing the adaptive correction of the model to the field working conditions.

[0079] In order to make the implementation content more complete and clear, the detailed process of drying treatment by those skilled in the art is attached:

[0080] Hot air stage: After the equipment is started, the main heater and the hot air heater are turned on at the same time, raising temperature 1 to the set value (for example, 85°C) and running for 120 minutes (parameter: heating time interval). The main heater temperature 1 and the hot air heater temperature 1 are simultaneously transferred to the main heater temperature 2 and the hot air heater temperature 2 for continuous heating. When the core temperature 1 reaches the set value (for example, 40°C), water is drained through the condenser. When the total running time of the hot air stage reaches the set value (for example, at least 360 minutes, including the set time of the heating time interval), and the core temperature 2 reaches the set value (for example, 60 degrees Celsius), the system enters the preheating stage;

[0081] Preheating and pressure changing stage: the hot air circulation is automatically closed, the vacuum degree is pumped from 80000Pa (inflating vacuum degree) to 40000Pa (evacuation vacuum degree), maintained for 5 minutes (evacuation and pressure holding time), and then inflated to 80000Pa, maintained for 5 minutes (inflating and pressure holding time), and then pumped from 80000Pa to 40000Pa, and the cycle is repeated. When the core temperature reaches 82℃ (parameter: core temperature setting) and the total running time reaches 180min (parameter: total running time), the vacuum degree is detected to see if it has reached 80000Pa. If not, it will be inflated to 80000Pa (inflating vacuum degree) and the preheating stage is over.

[0082] Step-by-step stage: pump from 80,000Pa to 40,000Pa, hold for 10 minutes (parameter: holding time setting), pump from 40,000Pa to 20,000Pa, hold for 10 minutes, pump from 20,000Pa to 10,000Pa, hold for 10 minutes and then enter the intermittent stage; during the holding process, regardless of the vacuum degree changes, only according to the time;

[0083] Intermittent stage: evacuate for 20 minutes, stop for 10 minutes, until the coil temperature reaches 100℃ (parameter: coil temperature) and the vacuum degree reaches 2000Pa (parameter: small Roots pump starting vacuum degree), start the small Roots pump, when the coil temperature reaches 100℃ and the vacuum degree reaches 1500Pa (parameter: large Roots pump starting vacuum degree), start the large Roots pump, when the evacuation vacuum degree reaches 400Pa (parameter: evacuation vacuum degree) and the coil temperature is above 100℃, enter the high vacuum stage;

[0084] High vacuum stage: Continue heating and vacuuming until the final dry vacuum reaches 50Pa (parameter: final dry vacuum), the coil temperature reaches 105℃ (parameter: final dry coil temperature) and the total operating time reaches 300min (parameter: total operating time). When these conditions are met at the same time, the endpoint judgment stage is carried out.

[0085] End point determination stage: When entering the end point determination stage, collect the vacuum degree once, wait for 2 minutes (parameter: vacuum determination interval time), collect the vacuum degree again, subtract the second vacuum degree from the first vacuum degree, and determine that the vacuum difference is no more than 20Pa. At the same time, the dew point temperature reaches -40℃, and the determination ends;

[0086] When the vacuum difference or dew point temperature is judged as unqualified, re-run high vacuum for 30 minutes (parameter: repeated judgment interval time), collect vacuum degree once at the end of 30 minutes, wait for 2 minutes (parameter: vacuum judgment interval time) and collect vacuum degree again, subtract the first vacuum degree from the second vacuum degree, and judge that the vacuum difference is not greater than 20Pa. At the same time, the dew point temperature reaches -40℃, and the judgment ends. And so on.

[0087] The device has no limit on the maximum number of times of judgment. As long as the standard is not met, the cycle will continue and the automatic mode can be stopped manually.

[0088] The above are merely embodiments of the present invention and are not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A transformer drying data acquisition and processing system includes a multi-source acquisition and processing module that communicates with all sensors during the drying process to collect five-dimensional dynamic data, specifically temperature, vacuum level, dew point temperature, vibration intensity, and instantaneous power. Furthermore, the multi-source acquisition and processing module stores static process data from the drying process, specifically coil resistance, insulation layer thickness, magnetic permeability, gap width between laminations, and ambient temperature and humidity. It is characterized in that Also includes: The prediction model building module, including the attention recognition unit, the temporal feature capture unit, and the fully connected unit, is used to build a drying endpoint prediction model based on the preprocessed data. Specifically: The attention recognition unit uses multi-head attention to automatically filter out the most critical time nodes and fluctuation patterns for determining the drying endpoint from the five-dimensional time series input and outputs a sequence code; Capturing time series feature units, using multi-layer long short-term memory networks based on sequence encoding to capture long-term dependencies in time series features to control information retention and updating; The fully connected unit extracts a vector representing the evolution of the entire process from the last hidden state of the top-level LSTM and concatenates it with the vector s of static process parameters, including coil resistance, insulation thickness, magnetic permeability, lamination gap, and average ambient humidity, in the feature dimension to form a fused vector z. After passing through a multi-layer fully connected network and activation, it ultimately outputs three regression indicators: remaining drying time, estimated energy consumption, and drying quality score. A multi-objective solution module is used to minimize the total drying time and minimize the cumulative energy consumption as the dual objectives and embed the physical constraints of dew point, vacuum change rate and temperature upper limit, and use a multi-objective evolutionary algorithm to generate a Pareto frontier solution; A real-time correction feedback module is used to fuse the model prediction and the actual completion time residual through a one-dimensional Kalman filter. When the residual exceeds the threshold, a local fine-tuning mechanism is triggered to dynamically correct the control curve and close the loop. The specific process of the multi-objective evolutionary algorithm is as follows: 3-1, extract the temperature curve D(t) and vacuum curve P(t), and express the two curves in the form of discrete control nodes. The composite objective function aims to minimize the total drying time and total energy consumption at the same time. The composite objective function is constructed and its expression is: ,in The total time required to complete one drying process is The accumulated power consumption to complete one drying process; The constraints of the composite multi-objective function are:

1. Dew point constraint: at the drying end point, the dew point temperature does not exceed -40°C; 2. Vacuum change rate constraint: the vacuum change rate does not exceed 20 Pa per minute; 3. Temperature upper limit constraint: the temperature does not exceed 110°C; 3-2, randomly generate several groups of candidate solutions, each group consists of the same number of control nodes, to ensure that all candidate initial solutions meet the constraints; each individual in the population explores the optimal balance through different curve combinations; for each candidate solution, call the fast simulation of the drying endpoint prediction model constructed above to predict when the dew point will reach -40℃ for the first time under the control of the temperature curve, and predict the total drying time required within the cycle as , and predict the cumulative energy consumption during the period If the temperature or vacuum rate is not constrained during the simulation, a penalty is imposed on the candidate solution, thereby obtaining the dual objective value of each individual , where n is a positive integer, representing the index of the candidate solution; 3-3, find the optimal control strategy based on the composite objective function and constraints of step 3-1.

2. The transformer drying data collection and processing system according to claim 1, characterized in that: The multi-source acquisition and processing module is also used to pre-process the raw data: Timing synchronization: Using whole seconds as the time unit, first downsample the high-frequency signals of temperature and vibration to once per second; then perform linear interpolation on the low-frequency signal of dew point to align all signals to the same time point; Outlier point removal: Calculate the historical mean and standard deviation of each time series. If the reading at a certain moment deviates from the mean by more than three times the standard deviation, the point is marked as an anomaly and removed to prevent instantaneous interference from affecting subsequent analysis. Real-time feature construction: At each timestamp, the first-order difference of physical quantities and the gradient change of dew point in space are calculated. This allows the original parameters to be time-series aligned and abnormal data to be eliminated to obtain clearly organized data, including temperature, temperature change rate, vacuum degree, vacuum change rate, dew point, dew point gradient, vibration intensity, and instantaneous power. Data standardization: To eliminate differences in the volume of dry tanks and the specific heat capacity of insulating materials across different factories or workshops, each time series feature is first scaled according to the inverse of the actual volume of the tank, and then adjusted a second time according to the thermal melting coefficient of the material. This allows all features to be mapped to the same standard volume and standard specific heat capacity, ensuring that the model inputs are dimensionally consistent and directly comparable.

3. The transformer drying data collection and processing system according to claim 2, characterized in that: Sequence encoding output process: Let the time series input matrix be , where T is the total number of time steps, d is the feature dimension of each step, and three sets of trainable linear mappings are projected onto the query, key, and value subspaces to generate query matrices, key matrices, and value matrices, respectively, where each row corresponds to a time step; for the query component at any time i and the key vector at all times j, the similarity score is calculated by dot product and divided by the scaling factor to measure the similarity between the state at the current time i and the historical time j; all scores are normalized to attention weights using the softmax function under the same query i , is the attention weight of the information at moment i to the information at moment j; For each moment i, all value vectors are weighted by the corresponding attention weights and summed to obtain the attention output, which represents the attention output vector at the i-th second, which is the weighted sum of all value vectors; In order to capture multiple attention patterns, h sets of parallel heads are actually used, each learning an independent query weight matrix, key weight matrix and value weight matrix. Finally, the outputs of the h heads are spliced ​​in the feature dimension and linearly mapped again to become an attention encoding sequence of unified dimension. ,in .

4. A transformer drying data collection and processing system according to claim 3, characterized in that: The process of capturing long-term dependencies in time series features is: In obtaining the attention encoding sequence After that, it is input into the multi-layer long short-term memory network to further extract the dynamic evolution characteristics across the time window. For each time step t, the multi-layer long short-term memory network maintains a cell state and a hidden state, and uses the forget gate and input gate to The output gate is used to control the retention and update of information. Since the cell state can be transmitted without attenuation across multiple time steps, important historical information can continue to affect the network output in the later stage of the sequence through the forget gate and input gate, thereby capturing the long-term dependencies of the drying process.

5. The transformer drying data collection and processing system according to claim 4, characterized in that: The process of the fully connected network outputting three regression results is: After the time series modeling is completed, the hidden vector at the last moment is extracted from the hidden state sequence of the top-level LSTM. This vector aggregates the dynamic evolution information from t=1 to T. At the same time, the static process parameters form a static feature vector. By concatenating the hidden vector and the static feature vector in the feature dimension, the final fusion vector z is obtained. This fusion vector then enters a multi-layer fully connected network. The last layer outputs a vector y=[y1,y2,y3] with a length of 3, where y1, y2, and y3 correspond to the remaining drying time, expected energy consumption, and drying quality score, respectively. To ensure physical rationality, the model imposes boundary constraints on the output: y1≥0, y2≥0, and y3∈[0,1].

6. The transformer drying data collection and processing system according to claim 5, characterized in that: The process of finding the optimal control strategy is: The population is stratified, and the first layer is the Pareto frontier: any solution in this layer has a and total energy consumption Neither of the two objectives can be dominated by other solutions in the same generation at the same time; the second layer is composed of solutions that are dominated by the first layer solutions but do not have a complete domination relationship with each other within the second layer; and so on, the entire population is divided into multiple non-dominated layers; Within each layer, the crowding degree of each solution in the two-objective space is calculated, that is, the distance between the solution and its neighboring solutions on the objective plane. The specific strategy is:

1. Prioritize the retention of non-dominated layers closer to the front; 2. Within the same layer, prioritize the retention of individuals with high crowding degrees. Next, a pairwise crossover is used from the retained pool to exchange some of the control nodes of the parent solution to generate a new generation of candidate solutions, thereby inheriting and mixing the outstanding time period control strategies of the parent. The above process is iterated continuously, and the drying time, energy consumption and quality penalty are recalculated in each generation. The new solutions are sorted by non-domination and screened by congestion until the Pareto front is stable or the preset maximum number of iterations is reached. When the algorithm converges, a set of mutually non-dominated optimal solutions is obtained, and each solution corresponds to a complete temperature-vacuum control curve and its drying time and energy consumption. Balanced results; according to the control scheme selected by the user or production scheduler, the corresponding node set values ​​are sent to the heater and vacuum pumping system through the interface in real time to form a closed-loop automatic adjustment.

7. A method for collecting and processing transformer drying data, characterized in that: The transformer drying data collection and processing system according to any one of claims 1 to 6 comprises the following steps: Step 1: Communicate with all sensors in the drying process to collect five-dimensional dynamic data, including temperature, vacuum level, dew point temperature, vibration intensity, and instantaneous power. In addition, the multi-source acquisition and processing module also stores static process data of the drying process, including coil resistance, insulation layer thickness, magnetic permeability, gap width between laminations, and ambient temperature and humidity. Step 2: Extract time series dynamic data and static process data to build key prediction models; Step 3: Based on the dual objectives of minimizing the total drying time and minimizing the cumulative energy consumption, a composite objective function that takes into account both time efficiency and energy consumption minimization is established, and a multi-objective evolutionary algorithm is used to solve it; In step 4, the model prediction and the actual completion time residual are fused through a one-dimensional Kalman filter, and when the residual exceeds the threshold, the local fine-tuning mechanism is triggered to dynamically correct the control curve and close the loop.

Images