A method for mapping multiple energy modes of an oven to cooking parameters

Abstract

This invention relates to a mapping method between multiple energy modes of an oven and cooking parameters. Addressing the problems of complex energy input characteristics, limited model generalization ability, and difficulty in achieving fine-grained adaptive control in traditional cooking processes, this invention proposes to acquire multi-dimensional temporal physical characteristics of multiple energy inputs, normalize them to generate uniquely identifiable energy fingerprints, and combine this with recipe information input into a lightweight master track model to output matching cooking parameters in real time. The system collects actual operating data and performs alignment and deviation identification with model predictions, triggering local model fine-tuning and energy fingerprint-incremental patch mapping updates to achieve adaptive correction of the model to new operating conditions. For multi-energy mixed scenarios, a composite fingerprint is generated through power weighting, and combined with a multi-source model to achieve thermal hysteresis compensation and power conversion jitter suppression. This scheme improves the dynamic adaptability of cooking parameters and energy utilization efficiency, while also possessing high precision and scalability.

Description

Technical Field

[0001] This invention relates to the field of intelligent cooking equipment control and multi-energy adaptive optimization technology, and in particular to a method for mapping oven multi-energy modes and cooking parameters. Background Technology

[0002] Current intelligent cooking equipment, especially multi-energy ovens, has gradually evolved from traditional mechanical and single-energy models to composite intelligent systems integrating multiple energy inputs (such as electricity, gas, and steam) in terms of precise control and parameter adaptation technology. Existing mainstream solutions largely rely on static recipe tables, energy consumption experience curves, rule-based scheduling based on knowledge graphs or expert systems, or parameter templates generated under different operating conditions through digital twin simulation. These methods generally use preset energy type labels (such as "electric mode" and "gas mode") to adapt a fixed set of preheating, temperature control, and heating mode parameters for each energy type, and combine the user-selected recipe or ingredient information with the energy mode label to achieve command output for the oven cavity temperature field, time segmentation, power setting, etc. Some high-end products have introduced adaptive closed-loop adjustment based on sensor feedback, but their parameter optimization process still relies on pre-trained models or offline big data analysis results, lacking real-time identification and targeted adaptation of the energy physical state.

[0003] With the increasing variety of energy types and combinations, the industry is also paying more attention to the generalization ability of parameter matching models under multi-energy synergy. For example, fluctuations in grid quality, differences in gas combustion characteristics, and thermal hysteresis and switching jitter under mixed energy modes can all cause significant disturbances in the actual cavity temperature rise rate, heat distribution uniformity, and energy consumption trajectory under the same recipe instructions, resulting in large fluctuations in cooking effects with the same parameter settings, or even cooking failure. Some manufacturers have attempted to add "energy sensors" for event-triggered corrections, but these usually only perform coarse-grained classification of energy types and do not delve into the multi-dimensional impact of energy transient characteristics and physical environment fluctuations. Furthermore, for multi-energy mixed input scenarios, existing technologies mostly rely on manual adjustment of weighting coefficients or preset compensation curves, lacking the ability to refine cooking parameters and adapt them in real time through self-learning under mixed energy inputs.

[0004] In the field of intelligent cooking equipment control and multi-energy adaptive optimization, typical application scenarios include automatic recipe scheduling in intelligent ovens under complex energy supply environments and simultaneous heating of multiple energy sources in digital kitchen systems. These technical approaches mainly address parameter matching under single energy conditions. When dealing with uncertainties in energy input combinations, parameter model generalization failures often occur, leading to unstable user experiences. Some systems rely on cloud or external database calls to achieve energy scenario identification and parameter tuning, which not only increases system complexity and communication latency but also makes it difficult to perform personalized local model iterative optimization.

[0005] Existing technologies face the following prominent problems: First, cooking parameter matching models generally only use energy category as a label, which cannot accurately describe the physical-layer temporal characteristics of energy, resulting in insufficient generalization of parameters such as energy efficiency and thermal inertia under the same "electric" or "gas" category; Second, under complex conditions with varying energy input combinations, there is a lack of a mechanism to identify the instantaneous physical state of energy in real time and dynamically generate parameters accordingly. When encountering energy switching, mixed operating conditions, or abnormal energy quality, the system cannot adaptively correct parameters in a timely manner, and the cooking effect is prone to fluctuations or even loss of control; Third, traditional parameter mapping models cannot achieve self-learning incremental adjustment based on historical usage trajectories, which prevents the equipment from becoming more accurate with repeated or similar scenarios, limiting the long-term evolution capability of intelligent cooking systems.

[0006] Therefore, there is an urgent need for a generalized solution that transcends traditional energy labeling, directly basing its approach on the real physical characteristics of energy inputs. Using high-dimensional energy fingerprints as a bridge, and through dynamic model calibration and localized incremental patching mechanisms, it can achieve real-time adaptive matching and updating of cooking parameters under multiple energy input scenarios. This ensures the stability of cooking effects, control robustness, and intelligent evolution capabilities of the system in complex energy environments. This technological breakthrough will greatly enhance the adaptability and user experience of intelligent cooking equipment, driving the industry towards a higher level of energy-parameter dynamic mapping and self-optimizing control. Summary of the Invention

[0007] This application provides a method for mapping multiple energy modes of an oven to cooking parameters, aiming to solve one of the problems or issues of the prior art mentioned in the background section.

[0008] This application provides a method for mapping multiple energy modes of an oven to cooking parameters, specifically including: S1: Acquire timing characteristic signals such as voltage fluctuation spectrum distribution, power ramp slope envelope, start-stop transient response delay, and thermal inertia decay time constant under multiple energy input conditions.

[0009] S2: Normalize the time-series feature signal to generate a high-dimensional vector as an energy fingerprint.

[0010] S3: Input the energy fingerprint and the user-selected recipe identifier into the main track model to output a set of cooking parameters.

[0011] S4: Collect the cavity temperature field curve, thermal image sequence and energy consumption trajectory after the control is executed, and compare the above real operation data with the cooking parameter set in time and space to generate deviation identification results.

[0012] S5: Based on the deviation recognition result, determine whether the prediction error under the same type of energy fingerprint input for a consecutive preset number of times exceeds the set threshold. If it does, trigger the fine-tuning command.

[0013] S6: In response to the fine-tuning instruction, freeze the subset of neurons in the main track model that are strongly correlated with specific parameters, and perform single-step gradient updates using the current batch running data to generate incremental patches.

[0014] S7: Establish an index mapping relationship between fingerprints and patches, and store the incremental patches in the calibration cache so that they can be automatically loaded and superimposed on the main rail output when the same energy fingerprint appears in the future, forming an adaptively corrected combination of cooking parameters.

[0015] S8: For multi-energy mixed input scenarios, the energy fingerprints of multiple independent energy sources are weighted and fused according to the real-time power ratio to generate a composite fingerprint, and the final control command is output by combining the main track model with thermal hysteresis compensation and power switching jitter suppression algorithm.

[0016] The mapping method between multiple energy modes of an oven and cooking parameters provided in this application has the following beneficial effects: (1) By introducing an energy fingerprint extraction unit and a multi-dimensional time-series feature sampling mechanism, this scheme significantly improves the perception accuracy and discrimination ability of different energy input characteristics, effectively overcoming the technical defects of poor generalization and weak adaptability of control strategies caused by traditional methods that rely solely on static category labels (such as "electricity" or "gas"). Since the energy fingerprint vector integrates physical layer dynamic features such as voltage / pressure fluctuation spectrum, power ramp slope envelope, start-stop transient response delay and thermal inertia decay time constant, and forms a high-dimensional vector representation with hash-like characteristics after normalization, even when the energy type is the same but the actual energy supply quality is different (such as voltage distortion under different grid loads and different gas source pressure fluctuation characteristics), the system can still accurately identify and distinguish its influence, thereby providing a more discriminative input basis for subsequent parameter matching, and greatly improving the relevance and rationality of the initial parameter setting in the cooking process.

[0017] (2) By constructing a dual-track mapping engine and a local incremental calibration mechanism, this scheme realizes the online self-evolution capability of the cooking control model, effectively solving the long-term performance degradation problem caused by environmental drift, equipment aging, or energy fluctuations in the existing technology. The lightweight main track model ensures real-time response capability, while the auxiliary track incremental calibrator performs spatiotemporal alignment analysis on the actual feedback of cavity temperature field evolution, thermal imaging sequence, and energy consumption trajectory to accurately locate the parameter dimensions and time periods with significant prediction deviations. When a certain type of parameter under a specific energy fingerprint is detected to continuously exceed the threshold error, the system only performs local fine-tuning on the relevant subset of neurons and stores the weight update in the form of incremental patches, avoiding the computational overhead and stability risks caused by retraining the entire model. Combined with the design of the fingerprint-patch index table, the system can automatically load historical optimization experience when the same energy conditions are reproduced, achieving a personalized adaptation effect of "the more it is used, the more accurate it becomes", which significantly enhances the robustness and long-term consistency of the control strategy.

[0018] (3) For multi-energy mixed input scenarios, this solution innovatively proposes a composite fingerprint fusion mechanism and a pre-trained master track model, which fully supports refined control under complex working conditions such as gas main heating + electric auxiliary heat preservation, breaking through the technical bottleneck of the traditional single-energy assumption that is difficult to deal with energy coupling effects. By generating composite fingerprints by weighting and fusing the fingerprint vectors of multiple energy sources according to the real-time power ratio, and calling a specially trained collaborative mapping sub-model, the system can actively compensate for thermal hysteresis, suppress power switching jitter, and coordinate the output rhythm of multiple sources, ensuring smooth temperature control transition and optimal energy efficiency allocation. The entire processing flow does not rely on external databases, knowledge graph scheduling or digital twin simulation platforms. All calculations are completed within the edge controller, which has the advantages of low latency, high reliability and network-free operation, and fully meets the requirements of embedded ovens for safety, real-time performance and deployment flexibility.

[0019] In summary, this solution reconstructs energy from traditional classification labels into measurable, memorable, and evolving physical entities, constructing a closed-loop adaptive control system from feature extraction and dynamic matching to online evolution. Without increasing hardware costs, it significantly improves the control accuracy, adaptability, and long-term stability of multi-energy cooking equipment in complex real-world scenarios, while also possessing good interpretability and scalability. This provides a new technical path for the efficient, precise, and personalized operation of smart kitchen appliances in heterogeneous energy environments. Attached Figure Description

[0020] Figure 1 This is the main flowchart of a method for mapping multiple energy modes of an oven to cooking parameters.

[0021] Figure 2 This is a sub-flowchart of a method for mapping multiple energy modes of an oven to cooking parameters.

[0022] Figure 3 This is another sub-flowchart of a method for mapping multiple energy modes of an oven to cooking parameters. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0024] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0025] like Figure 1 As shown, this application provides a method for mapping multiple energy modes of an oven to cooking parameters, specifically including: S1: Acquire timing characteristic signals such as voltage fluctuation spectrum distribution, power ramp slope envelope, start-stop transient response delay, and thermal inertia decay time constant under multiple energy input conditions.

[0026] S2: Normalize the time-series feature signal to generate a high-dimensional vector as an energy fingerprint.

[0027] S3: Input the energy fingerprint and the user-selected recipe identifier into the main track model to output a set of cooking parameters.

[0028] S4: Collect the cavity temperature field curve, thermal image sequence and energy consumption trajectory after the control is executed, and compare the above real operation data with the cooking parameter set in time and space to generate deviation identification results.

[0029] S5: Based on the deviation recognition result, determine whether the prediction error under the same type of energy fingerprint input for a consecutive preset number of times exceeds the set threshold. If it does, trigger the fine-tuning command.

[0030] S6: In response to the fine-tuning instruction, freeze the subset of neurons in the main track model that are strongly correlated with specific parameters, and perform single-step gradient updates using the current batch running data to generate incremental patches.

[0031] S7: Establish an index mapping relationship between fingerprints and patches, and store the incremental patches in the calibration cache so that they can be automatically loaded and superimposed on the main rail output when the same energy fingerprint appears in the future, forming an adaptively corrected combination of cooking parameters.

[0032] S8: For multi-energy mixed input scenarios, the energy fingerprints of multiple independent energy sources are weighted and fused according to the real-time power ratio to generate a composite fingerprint, and the final control command is output by combining the main track model with thermal hysteresis compensation and power switching jitter suppression algorithm.

[0033] Step S1: Acquire timing characteristic signals such as voltage fluctuation spectrum distribution, power ramp-up slope envelope, start-stop transient response delay, and thermal inertia decay time constant under multi-energy input conditions. Specifically, this includes: S1.1: Based on the embedded high-speed data acquisition interface, the original voltage and current timing signals of multiple energy input terminals are acquired. The original voltage timing signals are then processed by frequency domain decomposition using the fast Fourier transform algorithm to extract the voltage fluctuation spectrum distribution data.

[0034] Based on an embedded high-speed data acquisition interface, the raw time-series signals output from voltage and current sensors at multiple energy input terminals are synchronously acquired to ensure a precise correspondence between the two signals under the same time reference. A window function is applied to the acquired voltage time-series signal for preprocessing, suppressing spectral leakage and enhancing frequency resolution within a specified sampling length, creating stable conditions for subsequent frequency domain decomposition. The Fast Fourier Transform (FFT) algorithm is then used to map the windowed voltage time-series signal from the time domain to the frequency domain, generating a spectrum vector containing the amplitude and phase of each frequency component. The amplitude spectrum is calculated based on the spectrum vector, and the frequency energy distribution within a specific bandwidth is extracted as the voltage fluctuation spectrum distribution data. The energy calculation formula is as follows: in, Energy value For frequency component index, The upper limit sampling point number for the specified frequency band. For the first The amplitude of each frequency component is calculated. Feature vectors are constructed from the spectral characteristics under different energy conditions, and a set of spectral parameters, including grid quality indicators and gas ignition interference characteristics, is labeled and stored as voltage fluctuation spectral distribution data, which is then output to the subsequent slope calculation step. Through this processing method, the synchronously acquired signal from the previous step is transformed into structured voltage fluctuation spectral distribution data, achieving an accurate frequency domain characterization of the transient characteristics of energy input.

[0035] For example, in a household multi-energy oven test, the embedded acquisition module synchronously acquires voltage and current signals at a sampling rate of 10000Hz, with a sampling window length of 2048 points. The voltage signal is preprocessed using a Hanning window function. An FFT is performed to obtain 1024 effective frequency components. Energy integration is performed on the amplitude spectrum within the 0–500Hz range, calculating the grid quality energy value to be 3.25 volts square units. In gas heating mode, specific 50Hz harmonics and 150Hz interference peaks appear in the spectrum, with an energy value of 2.87 volts square units, significantly different from the spectrum distribution in pure electric heating mode. This energy characteristic spectrum vector is passed to the power ramp-up slope calculation module to verify that the spectrum pattern under different energy inputs can stably reflect the transient characteristics of energy, improving the accuracy and robustness of the subsequent parameter matching model in identifying energy types.

[0036] S1.2: Based on the time-domain window corresponding to the voltage fluctuation spectrum distribution data, the slope of the original power sequence is calculated using a sliding differential filtering algorithm to extract the power ramp-up slope envelope data that reflects the thermal response speed of the heating element or the ignition efficiency of the burner.

[0037] Based on the time-domain window corresponding to the voltage fluctuation spectrum distribution data obtained in step S1.1, the power calculation module is invoked to perform point-by-point multiplication on the original current time-series signal within the same window, generating an original power sequence with time as the horizontal axis and instantaneous power as the vertical axis. A sliding differential filtering algorithm is used to perform differential processing on the original power sequence. Under the condition that the sliding window length matches the sampling frequency, the power change rate at each time point is calculated, and a slope sequence containing the slope values ​​of all time points is output. The slope sequence is subjected to amplitude envelope extraction. A peak tracker is invoked to identify significant positive slope segments in the power ramp-up phase while maintaining the time order, and an envelope curve is generated to characterize the upward trend of the slope over time. By fitting the envelope curve and the physical model of the power ramp-up phase, the heating element thermal response speed index or burner ignition efficiency index is calculated to evaluate the dynamic performance of the energy input end within this time-domain window. The aforementioned envelope curve and performance index are encapsulated as power ramp-up slope envelope data, serving as the input source for subsequent start-stop response delay analysis. By using sliding differential filtering and envelope extraction, the voltage spectrum time-domain window result from the previous step is transformed into power change trend data with physical meaning, thereby achieving accurate quantification of the transient response characteristics of heating components under multi-energy conditions.

[0038] For example, under the electrical energy operation of a multi-energy hybrid heating oven, the sampling frequency is configured to 200Hz, the time-domain window length is 5 seconds, and the original power sequence is obtained by multiplying the instantaneous voltage and current. In the sliding differential filter, the window length is set to 10 sampling points, and the difference operation formula is as follows: in, The slope for Instantaneous power value at any given moment The time sampling point index is used. Envelope extraction was performed on the slope sequence, detecting a maximum positive slope of 3.2 W / ms during the ramp-up phase, lasting 1.8 seconds. The integral value of the fitted envelope curve was used to calculate the thermal response rate, which was 5.76 kW / s. This value is significantly higher than the natural heating rate of similar equipment, indicating that the heating element under this energy input has a significantly improved transient heating performance and can provide a highly reliable performance benchmark input for subsequent start-stop delay analysis.

[0039] S1.3: Based on the starting trigger point of the power ramp slope envelope data, the timing alignment analysis of the energy start-stop control command and the actual load response waveform is performed using zero-crossing detection and threshold comparison logic to extract start-stop transient response delay data.

[0040] Based on the starting trigger point of the power ramp slope envelope data, the embedded event marking module is invoked to synchronously initialize the time axis of the energy start-stop control command signal and the actual load response waveform, ensuring that the two types of signals have a unified sampling start time. Zero-crossing detection is performed on the energy start-stop control command signal, generating a set of command trigger event timestamps by identifying zero-crossing positions in the control signal waveform where the signal level changes from low to high or from high to low. Threshold comparison logic is applied to the actual load response waveform, continuously comparing the waveform amplitude with a preset load response start threshold. When the waveform amplitude first exceeds the threshold, the physical response start timestamp is recorded. Timing alignment analysis is performed based on the command trigger event timestamp and the physical response start timestamp, calculating the time difference between the two to quantify the response hysteresis characteristics of the actuator. The start-stop transient response delay is determined using the following delay calculation formula: in, For the load response start timestamp, To control the trigger timestamp of the command, This represents the transient response delay. The calculated transient response delay is mapped to the mechanical delay of the actuator or the hysteresis of the solenoid valve, providing delay feature components for subsequent thermal inertia decay model input.

[0041] By using zero-crossing detection and threshold comparison, the starting point of the power ramp slope envelope from the previous step is transformed into precise energy control-load timing alignment delay data, thereby achieving high-precision calibration of delay characteristics in energy fingerprints.

[0042] For example, for a commercial oven using gas as the main heater and electric as the auxiliary heater for insulation, the sampling frequency of the energy start / stop control command signal is 1000Hz, the sampling frequency of the actual load response waveform is 500Hz, the starting time point of the power ramp-up slope envelope is located at 250ms, the zero-crossing detection of the control command waveform determines 260ms as the command trigger time stamp, and the threshold comparison of the load response waveform determines 420ms as the physical response start time stamp. Substituting the data into the delay calculation formula, we obtain... 0.160 seconds indicates that the mechanical delay of the oven's gas actuator is 160 milliseconds. Under five consecutive inputs of the same energy fingerprint, this delay value fluctuates by no more than 5 milliseconds, verifying the stability and repeatability of the delay feature calibration and providing highly reliable energy transient response characteristics for subsequent steps.

[0043] S1.4: Based on the cooling stage after the start-stop transient response delay data ends, the attenuation model matching process of the cavity temperature drop curve is performed using an exponential fitting regression algorithm to extract the thermal inertia decay time constant data.

[0044] After the cooling phase following the start-stop transient response delay data, the high-precision thermocouple or infrared temperature sensor installed in the oven cavity outputs the cavity temperature timing signal, and constructs a temperature drop data sequence based on this signal according to a fixed sampling period.

[0045] Noise suppression processing is performed on the temperature decrease data sequence. The Savitzky-Golay smoothing filter algorithm is used to eliminate high-frequency disturbances while preserving the curve trend, generating a smoothed temperature curve.

[0046] Identify the start and end times of the cooling phase on the smoothed temperature curve, establish a truncation window to extract temperature time segments containing only the cooling process, and ensure that the attenuation model matches the input range accurately to correspond to the physical cooling phase.

[0047] An exponential decay regression model is constructed based on the temperature time series segment of the cooling stage. The time constant parameter of the waste heat release process is estimated using the least squares fitting method, wherein the temperature decay formula is defined as: in for The cavity temperature at any given time, For ambient temperature, This is the starting temperature of the cooling phase. This is the thermal inertia decay time constant.

[0048] By analyzing the formula above Regression fitting calculations are performed to minimize the residuals between the actual data of the cooling curve and the theoretical model, and the time constant index characterizing the furnace insulation performance and the waste heat release characteristics of the heat source is output.

[0049] The thermal inertia decay time constant data obtained from regression is appended to the time series feature signal set to achieve a quantitative description of the cooling stage in energy fingerprint generation.

[0050] By using exponential fitting regression, the temperature data of the cooling stage in the previous step is transformed into a thermal inertia decay time constant, thereby enabling precise quantification of the residual heat release characteristics and heat preservation performance of the oven under different energy input conditions.

[0051] For example, in a commercial multi-energy oven, when both gas main heating and electric auxiliary heating modes are simultaneously activated, after the start-stop transient response delay is detected, the cavity temperature drops from 250°C to room temperature (25°C). The sampling period is 1 second, and the total acquisition time is 900 seconds. The temperature sequence is processed using Savitzky-Golay filtering with a window length of 51 points and a polynomial order of 3 to obtain a smooth curve. A temperature segment during the cooling phase is extracted, with an initial temperature of 250°C and an ambient temperature of 25°C. Least square fitting is performed using the aforementioned exponential decay model, yielding a thermal inertia decay time constant of 420 seconds. This result is stored as the thermal inertia decay time constant in the feature set. In the energy fingerprint generation stage, this value, along with features such as voltage fluctuation spectrum, power ramp-up slope, and start-stop transient response delay, constitutes a uniquely identifiable input vector. Model verification shows that this constant can significantly improve the stability of the cooking curve and reduce temperature overshoot when switching energy types.

[0052] S1.5: Based on the voltage fluctuation spectrum distribution data, power ramp slope envelope data, start-stop transient response delay data, and thermal inertia decay time constant data, the above-mentioned multidimensional heterogeneous physical quantities are aggregated into the time-series characteristic signal in a unified format using a multi-channel data fusion encapsulation protocol, which serves as the standard input object for subsequent normalization processing.

[0053] Step S2: Normalize the time-series feature signal to generate a high-dimensional vector as the energy fingerprint. Specifically, this includes: S2.1: Obtain time-series characteristic signals such as voltage fluctuation spectrum distribution, power ramp-up slope envelope, start-stop transient response delay, and thermal inertia decay time constant. Perform dimensionless processing on the time-series characteristic signals based on the range normalization algorithm to generate a set of standardized characteristic components that eliminates the influence of dimensions.

[0054] It should be noted that the main track model is a lightweight, fully connected network deployed within the oven's edge controller. Its input is a fusion tensor of the energy fingerprint vector and recipe identifier, and its output is a set of basic cooking parameters (heating power, damper opening, and time segmentation). The main body of the model consists of three hidden layers (fully connected layer + Bi-GRU layer + attention layer), and the output layer contains three parallel branches. The model supports local fine-tuning: single-step gradient updates are performed only on a subset of neurons strongly correlated with prediction bias, generating incremental patches that are then superimposed on the main track output. For multi-energy mixed input scenarios, the model expands to a collaborative branch, receiving a power-weighted composite fingerprint and outputting parameters for thermal hysteresis compensation and power switching jitter suppression.

[0055] S2.2: Receive the standardized feature component set, and perform a spatial mapping transformation on the standardized feature component set based on the weighted Euclidean distance metric method to generate an initial high-dimensional feature vector characterizing the transient characteristics of energy input.

[0056] Using the standardized feature component set output from the preceding steps as the input data object, the weighted Euclidean distance metric module is invoked to read the numerical range and weight coefficients of each dimension's feature components. Each feature component is then multiplied element-wise according to its preset weight coefficients to generate a weighted feature component set, providing a numerical basis for subsequent spatial mapping transformations. Euclidean distance calculation is performed on the weighted feature component set, mapping the distance metric values ​​to a multidimensional spatial coordinate system. A linear mapping matrix is ​​used to convert each weighted component into high-dimensional spatial point coordinates, forming the coordinate components of the initial high-dimensional feature vector. The mapped spatial coordinate set is then normalized and re-mapped to ensure dimensional consistency across different dimensions and to match the contribution ratio of each feature component to the transient characteristics of the energy. The output initial high-dimensional feature vector serves as a unified physical index for the transient characteristics of the energy input in the parameter matching module, realizing the vectorized expression of energy fingerprint features.

[0057] By using a weighted Euclidean distance metric and a multidimensional space mapping process, the standardized feature component set from the previous step is transformed into an initial high-dimensional feature vector that can comprehensively characterize the transient characteristics of energy input, thereby achieving accuracy and scalability of energy fingerprint vectorization.

[0058] S2.3: Using the initial high-dimensional feature vector, the initial high-dimensional feature vector is subjected to dimensionality reduction and discretization encoding based on the local sensitive hash projection mechanism to generate a binary fingerprint code sequence with hash collision-like characteristics.

[0059] It receives the initial high-dimensional feature vector as input, calls the parameter initialization module of the locality-sensitive hash projection mechanism to set the random vector distribution of the projection matrix and the projection threshold set, and ensures that the projection process has the ability to preserve the local neighborhood structure of the feature space.

[0060] The initial high-dimensional feature vector is multiplied by the projection matrix to obtain the intermediate feature vector after projection. This operation is implemented on the embedded controller by a multi-threaded floating-point unit to reduce computational latency.

[0061] The intermediate feature vectors after projection are compared with the preset threshold set. The comparison results are mapped to binary states through a sign function to form binary encoded segments on independent projection axes.

[0062] The binary encoded segments on each projection axis are concatenated and combined in a predefined index order to form a binary fingerprint code sequence with locality-sensitive hash collision characteristics. This sequence structure can generate a high probability of hash key value consistency when energy input characteristics are similar.

[0063] Collision optimization logic is introduced to detect the repetition rate of the generated binary fingerprint code sequence in the historical fingerprint set. Bitwise XOR perturbation is performed on bit segments with excessive repetition rate to enhance the distinguishability of energy fingerprints and reduce the probability of misjudgment.

[0064] By using the locality-sensitive hashing projection mechanism, the initial high-dimensional feature vector from the previous step is transformed into a binary fingerprint sequence that has stable collision characteristics in the energy feature space and can be used for subsequent fixed-length vector generation, thereby achieving efficient encoding and identification of multi-energy transient features.

[0065] For example, in a commercial multi-energy oven system, when the initial high-dimensional feature vector generated by weighted Euclidean mapping has a dimension of 128 and a value range of [-1, 1], the size of the local sensitive hash projection matrix is ​​set to 128 × 64, the random vector follows a normal distribution with a mean of 0 and a standard deviation of 0.5, and the threshold set is uniformly divided into zero values ​​according to each projection axis. When performing matrix multiplication, the SIMD instruction set of the embedded DSP is used to implement parallel computation of 64-axis projection to obtain a 64-dimensional intermediate feature vector. Each component of the intermediate feature vector is compared with the corresponding threshold. Those greater than the threshold are mapped to 1, and those less than or equal to the threshold are mapped to 0, forming a 64-bit binary encoded segment. The collision probability of the combined binary fingerprint code sequence in the historical energy fingerprint database is measured to be 12%, which is higher than the 5% threshold set by the system. This triggers the perturbation logic to perform an XOR perturbation on the bit segment index interval [16, 23]. After the perturbation, the collision probability decreases to 4%, significantly improving the distinguishing performance of the energy fingerprint. The final output binary fingerprint sequence is passed to step S2.4 to perform fixed-length vector generation, verifying that the fingerprint collision rate is consistently below 5% in the scenario of switching between electricity and gas, effectively increasing the accuracy of multi-energy mode mapping and cooking consistency.

[0066] S2.4: Based on the binary fingerprint code sequence, perform dimension alignment operation on the binary fingerprint code sequence according to a fixed-length padding and truncation strategy to generate a fixed-length energy fingerprint vector.

[0067] The binary fingerprint sequence generated in the preceding steps is received as the input object, and a processing link for generating a fixed-length vector is established to ensure the dimensional uniformity of the energy fingerprint.

[0068] The actual length of the binary fingerprint code sequence is calculated by difference from the preset length vector length. If the difference is positive, padding bits are added to the end of the sequence according to the fixed length padding strategy. The padding bits adopt an all-zero or all-one mode and a length check tag is placed in the additional bit segment to ensure the reversibility of the padded sequence.

[0069] When the difference is negative, the bits at the end of the sequence are truncated according to the fixed-length truncation strategy. Before truncating, the bit priority sorting algorithm is called to rearrange the bit order to retain high-weight feature bits. After truncating, a truncation flag is added to the end to prevent decoding errors.

[0070] By using bitwise operations and a block location mapping table, a dimension alignment operation is performed on the padded or truncated sequence to recombine each logic block into a continuous high-order bit to low-order bit layout, ensuring the consistency of the sequence in the hardware address mapping.

[0071] The vector constructor is called to assemble the dimensionally aligned bit data into a fixed-length energy fingerprint vector at a preset fixed length, and a cyclic redundancy check code is embedded during the assembly process to provide internal error detection capability.

[0072] By using fixed-length padding and truncation, the binary fingerprint code sequence from the previous step is transformed into a fixed-length energy fingerprint vector with fixed length, reversibility, and verification function, thereby achieving dimensional standardization and improved matching accuracy of the energy fingerprint input in subsequent model processing.

[0073] For example, a commercial multi-energy oven generates a 92-bit binary fingerprint sequence in gas main heating + electric auxiliary heat preservation mode. The preset long vector length is 128 bits, with a difference of 36 bits. The padding strategy uses an all-zero mode and adds an 8-bit check segment at the end. After padding, a bit priority sorting algorithm is used to retain the first 64 high-weight bits representing the gas thermal response characteristics, and the redundant bits of electric heating are placed at the end for priority removal in trimming scenarios. During dimension alignment, the fingerprint code is shifted and mapped to the sequential address space in 8-bit blocks. The final constructed 128-bit fixed-length energy fingerprint vector is embedded through cyclic redundancy check (CRC), with a check code length of 16 bits, and the detection result is zero errors. This vector can be directly verified in the subsequent integrity encapsulation in S2.5. At the same time, when the recipe identifier is input into the main track model in S3, the matching input of energy transient characteristics and dish requirement attributes is not affected by dimension inconsistency. The cooking parameter combination output by the main track model significantly improves stability in terms of thermal hysteresis compensation and power switching jitter suppression.

[0074] S2.5: Call the fixed-length energy fingerprint vector and perform integrity encapsulation processing on the fixed-length energy fingerprint vector based on checksum verification logic to generate a uniquely identifiable energy fingerprint that can be used for model input.

[0075] like Figure 2 As shown, step S3 involves inputting the energy fingerprint and the user-selected recipe identifier into the main track model to output a set of cooking parameters. Specifically, this includes: S3.1: Based on the energy fingerprint vector with hash-like properties generated in the previous steps and the recipe identifier selected by the user interface, perform multimodal data alignment and joint embedding processing to construct a joint input tensor containing transient physical features of energy and thermodynamic requirements of the dish.

[0076] Based on the fixed-length energy fingerprint vector with hash-like properties generated in the previous steps and the recipe identifier selected by the user interface, the multimodal data preprocessing unit loads the numerical values ​​of each dimension of the energy fingerprint vector and the set of thermodynamic feature parameters corresponding to the recipe identifier into a unified data cache. The energy fingerprint vector and the set of recipe thermodynamic feature parameters are mapped to the same numerical domain, and a normalized matrix transformation is used to ensure the comparability of features from different sources on a numerical scale. For each dimension of the energy fingerprint's physical transient features, weights are adjusted according to the set of heat sensitivity coefficients associated with the recipe identifier to generate an energy feature weighting matrix. The set of recipe thermodynamic feature parameters is arranged in a preset feature order, and a positional encoding mechanism is used to introduce the temporal correlation information of the recipe features, resulting in a recipe feature encoding matrix. A joint embedding operator is used to perform element-level fusion operations on the energy feature weighting matrix and the recipe feature encoding matrix to generate a multimodal joint feature matrix. The deep tensor construction interface is called to map the multimodal joint feature matrix into a joint input tensor acceptable to the model. The dimensional structure meets the input layer configuration requirements of the subsequent main track model. Through multimodal data alignment and joint embedding processing, the result of the previous step is transformed into a joint input tensor containing transient physical features of energy and thermodynamic requirements of dishes, thus realizing the preparation of input data for initial cooking parameter inference.

[0077] For example, in a multi-energy oven control system, the fixed-length energy fingerprint vector has 128 dimensions, with each dimension's value ranging from 0 to 1 after normalization. The set of thermodynamic feature parameters corresponding to the recipe identifier includes 24 features such as preheating temperature, target temperature, heating rate, and heat holding time, all of which are converted to the normalized domain. The energy feature weight matrix is ​​calculated using the recipe's heat sensitivity coefficient. For instance, for a recipe with a heating rate sensitivity coefficient of 0.8, the weight adjustment value for the corresponding heating-related feature dimension in the energy fingerprint is the original value multiplied by 0.8. The recipe feature encoding matrix uses a positional encoding vector. Generate, where For feature serial number, The angular frequency is encoded. The energy feature weighting matrix and the recipe feature encoding matrix are fused according to the formula. To merge, For energy weight matrix, For energy fingerprint vector matrix, The recipe encoding matrix outputs a joint embedding feature matrix with dimensions of 128×24. This matrix is ​​transformed into a 128×24×1 input tensor structure via a tensor construction interface. After being input into the main track model, it can significantly improve the matching accuracy of the target cooking curve under multi-energy conditions. Validation shows that the difference between the cooking parameter set generated under the condition of mixed gas and electricity supply and the actual needs is significantly reduced.

[0078] S3.2: Using the master track model pre-deployed in the edge controller, perform deep feature extraction and forward propagation calculation on the joint input tensor to obtain a high-dimensional hidden layer state feature map representing the matching degree between the current energy supply capacity and the target cooking curve.

[0079] Based on the joint input tensor constructed in the previous steps, the input interface of the main track model deployed in the edge controller is called to write the tensor data into the model input buffer to initialize the computation graph structure.

[0080] After initializing the computation graph, a multi-scale convolution operation is performed on the input tensor through a convolutional operation layer. The size of the convolution kernel is set in the spatial dimension and the temporal dimension according to the feature scale of the energy fingerprint vector and the thermodynamic feature requirements of the recipe identifier, respectively. Local correlation features are extracted to reflect the microscopic influence of energy transient features on the cooking process.

[0081] Following the convolution output, the batch normalization processing unit is called to normalize the convolution feature map, ensuring that the feature distribution after fusing different energy fingerprints and recipe characteristics presents a stable numerical scale within the model, reducing the internal covariate shift effect.

[0082] The normalized feature map is input into the bidirectional gated recurrent unit layer. The forward and reverse computation paths of the recurrent unit are used to capture the bidirectional dependence of energy features and recipe thermodynamic requirements in time, thereby outputting an implicit state sequence containing time context information.

[0083] The attention mechanism module is invoked to calculate the weight coefficients of each time slice in the hidden state sequence. Based on the weight coefficients, a weighted summation operation is performed on the hidden state sequence to enhance the ability to focus on the features of energy supply capacity and key periods of the cooking process.

[0084] The weighted summed vector is mapped to a high-dimensional hidden layer. The feature dimension is increased by combining a fully connected layer with a nonlinear activation function to form a high-dimensional hidden layer state feature map that represents the matching degree between the current energy supply capacity and the target cooking curve.

[0085] Through the aforementioned deep feature extraction and forward propagation computation processing, the joint input tensor of the previous step is transformed into high-dimensional hidden layer state features that can be used to decode cooking parameters, thereby realizing a dynamic matching measurement between energy and recipe requirements.

[0086] For example, in an embedded edge controller oven with dual energy inputs of electricity and gas, the energy fingerprint vector length is set to 128 dimensions, the recipe identifier length after embedding is 64 dimensions, and the joint input tensor size is 192×10×10 (channel×space×time). The convolutional kernel has a spatial dimension of 3×3 and a temporal dimension of 3, with 64 convolutional channels. The size of the local feature map extracted by convolution is 64×8×8, and batch normalization is applied to all convolutional channels. The hidden state dimension of the bidirectional gated recurrent unit is set to 128. After combining forward and backward paths, the output hidden state sequence length is 8, and the attention weight vector length is 8. Each weight coefficient is calculated using the following formula: in, This is the inner product of the hidden state and the attention parameters. The attention coefficients correspond to the time slices. The hidden states are then... Weighted summation yields a focused feature vector of length 128, which is then upscaled to 256 dimensions via a fully connected layer and activated by ReLU to generate a high-dimensional hidden state feature map. This feature map is used for cooking parameter decoding in subsequent steps. This process significantly improves the stability of cooking parameter prediction and maintains consistency in cooking curves under dynamic energy switching scenarios.

[0087] S3.3: Based on the high-dimensional hidden layer state feature map, the nonlinear activation function of the model output layer and the fully connected mapping operation are used to perform a decoding transformation from the abstract feature space to the specific control parameter space, so as to generate a cooking parameter set covering the heating power setting value, the damper opening adjustment amount and the time segment threshold.

[0088] Based on the input conditions of the high-dimensional hidden layer state feature map, a high-dimensional feature map generated by the second stage forward propagation of the main track model is received. The feature map contains multi-dimensional representation signals of energy transient physical properties and recipe thermodynamic matching attributes.

[0089] The high-dimensional hidden layer state feature map is input into the output fully connected layer of the model. Matrix multiplication is performed on the weights of each group of neurons to map the hidden layer features to the specific cooking control parameter space according to the arrangement of the output layer neurons.

[0090] In the fully connected layer output structure, a nonlinear activation function is applied to compress the amplitude or threshold the linear mapping result through functions such as ReLU, Sigmoid or Tanh, so that the output quantity conforms to the physical constraints of the cooking system control command.

[0091] For different cooking parameter value ranges, a normalized inverse mapping table is constructed based on the preset parameter sensitivity coefficient. The activated output value is then converted into the actual control parameter value through table lookup or interpolation.

[0092] The transformed parameter set is decomposed into a structured cooking parameter set containing heating power setting value, damper opening adjustment amount and time segment threshold, and boundary constraint verification is performed on each parameter to ensure that it is within the safe and process-permissible range.

[0093] By using fully connected mapping and nonlinear processing, the hidden layer feature output from the previous step is transformed into cooking parameter data that can directly drive the actuator, thus achieving a precise match between energy characteristics and cooking needs.

[0094] For example, in a commercial multi-energy oven, a high-dimensional hidden layer state feature map with a dimension of 128×1 is generated after a fixed-length energy fingerprint vector is input into the main rail model. This feature map is then input into the output layer, which has 3 neurons, corresponding to the heating power setpoint, damper opening adjustment, and time segmentation threshold, respectively. ReLU is used as the nonlinear activation function to process the output vector 2350, 60, 15, resulting in 2350, 60, 15. After parameter normalization and inverse mapping, these are converted into a heating power of 2350W, a damper opening of 60%, and a time segmentation threshold of 15 minutes, respectively. The formula for calculating the heating power setpoint is as follows: ,in For the corresponding neuron weights, The hidden layer state characteristic components; the formula for calculating the damper opening adjustment is: The formula for calculating the time segment threshold is: In actual verification, after the cooking parameter set drives the actuator, the cavity preheating stage is highly consistent with the target temperature curve, the opening and closing of the damper is stable, the segmented control switching is smooth, the overall performance is significantly improved, and it meets the requirements of commercial kitchens for cooking consistency in multi-energy environments.

[0095] like Figure 3 As shown, step S4 involves acquiring the cavity temperature field curve, thermal imaging sequence, and energy consumption trajectory after control execution, and then performing a spatiotemporal comparison between the aforementioned real operating data and the cooking parameter set to generate deviation identification results. Specifically, this includes: S4.1: Synchronously acquire the cavity temperature sensor sequence, food infrared thermal image frame sequence, and power metering pulse sequence after control execution, and use the multi-source sensor timestamp interpolation algorithm to generate cavity temperature field curve, thermal image sequence, and energy consumption trajectory with a unified time reference.

[0096] The cavity temperature sensor sequence, food infrared thermal image frame sequence, and power metering pulse sequence are synchronously acquired after control execution. The acquired objects include temperature timing signals output by thermocouples or thermistors distributed at different locations in the furnace, food surface thermal image frame sequences obtained by the infrared / visible light imaging module, and pulse counting sequences recorded by the power metering unit. A multi-source sensor timestamp interpolation algorithm is applied to the temperature timing signals to map the local clock records of each sampling node to a unified reference clock domain. The interpolation process includes calculating virtual sampling points according to the interpolation factor and filling in missing time slices. A timestamp consistency operation is performed on the thermal image frame sequence, matching the acquisition time information of the imaging frames with the unified reference clock domain of the temperature sequence. A thermal image lookup table consistent with the temperature sampling interval is generated using a bilinear time interpolation method. The pulse timestamp and reference clock domain synchronization processing is performed on the power metering pulse sequence. A cumulative count conversion function is introduced to map the pulse sequence into a continuous energy consumption numerical curve. This cumulative function can be expressed as: in Where C is the energy consumption, C is the number of pulse accumulations, and K is the energy constant corresponding to a single pulse. For the current unified timestamp, The starting timestamp is used. The above three types of data are fused from multiple sources according to a unified time base to form cavity temperature field curves, thermal image sequences and energy consumption trajectories with consistent time indexes. Through this multi-source time synchronization processing method, the execution results of the initial cooking parameters in the previous step are converted into standardized actual operation data that can be directly used for spatiotemporal comparison, thereby achieving consistency and high-precision matching of multi-source observation data.

[0097] For example, in a commercial oven with three energy input modes, the sampling frequency of the cavity temperature sensor is set to 2Hz, the frame rate of the thermal imaging sensor is set to 1Hz, and the pulse resolution of the energy metering unit is 1000 pulses per kilowatt-hour. During a certain cooking process, the original timestamps of the temperature sequence are at 0.5-second intervals, the timestamps of the thermal imaging frames are at 1-second intervals, and the recording time of the energy metering pulses is irregular. A multi-source sensor timestamp interpolation algorithm is applied, using 0.5 seconds as a unified time base. Linear interpolation is used to pad the thermal imaging frames to a 2Hz sampling rate, and a cumulative counting method is used to convert the energy pulses into an energy consumption curve. In the energy consumption curve calculation, assuming the cumulative number of pulses is 250, the energy constant corresponding to a single pulse is 1Wh, the current unified timestamp is 125 seconds, and the initial timestamp is 0 seconds, then the energy consumption value is 250Wh. After synchronization, the three types of data have corresponding values ​​at each 0.5-second sampling point. The temperature field curve shows that the heating behavior of the heat source gradually stabilizes. The thermal image frame confirms that the surface temperature and internal temperature gradient of the food tend to be uniform. The energy consumption curve provides a continuous record of energy utilization efficiency. Finally, the fused data output significantly improves the correlation accuracy and model judgment accuracy in subsequent deviation analysis.

[0098] S4.2: Based on the cavity temperature field curve, thermal image sequence and energy consumption trajectory, call the pre-stored spatiotemporal coordinate mapping matrix to map the multidimensional physical quantity data to the same discrete time slice and spatial grid as the cooking parameter set, so as to generate a spatiotemporally aligned actual operation data set.

[0099] Using cavity temperature field curves, thermal imaging sequences, and energy consumption trajectories based on a unified time reference as input objects, a pre-stored spatiotemporal coordinate mapping matrix is ​​invoked to provide the correspondence between multidimensional physical quantities and standard time slices and spatial grids.

[0100] Each set of temperature samples in the temperature field curve is repositioned to the same discrete time slice position as the cooking parameter set according to the time index of the mapping matrix, thus achieving standardized alignment of the time axis.

[0101] The pixel matrix in the infrared thermal image sequence is divided into regions and transformed into a gridded projection based on the spatial grid index provided in the mapping matrix, so as to convert the thermal distribution data in the original pixel coordinate system into thermal intensity values ​​on the standard spatial grid.

[0102] The energy consumption pulses or gas consumption data in the energy consumption trajectory are segmented and summed according to the time slice boundaries in the mapping matrix to make them completely correspond to the temperature field curve and thermal image sequence in the time dimension.

[0103] By adopting a multi-dimensional data structure index merging method, time-aligned temperature data, spatially aligned thermal image data, and segmented summed energy consumption data are merged into a composite physical quantity matrix, forming a spatiotemporally aligned actual operation data set containing three types of data: temperature, heat distribution, and energy consumption.

[0104] By standardizing the mapping matrix, the results of the previous step are transformed into a data set that is completely consistent with the cooking parameter set in terms of time and space, achieving precise spatiotemporal alignment with the actual operating state and providing a non-biased comparison basis for subsequent deviation feature extraction.

[0105] For example, in a hybrid energy oven with three-phase electric heating and gas-assisted insulation, the cavity temperature field curve is collected once per second by three sets of thermistors distributed at the top, middle, and bottom. The infrared thermal image sequence is collected by a 640×480 pixel thermal imager at 2Hz. The energy consumption trajectory is recorded by an energy meter every 0.5 seconds for the energy increment and by a gas flow meter every 1 second for the gas flow increment. The pre-stored mapping matrix sets the time slice interval to 1 second, and the spatial grid is divided into 8×8 regions. The temperature data of the temperature field curve is repositioned to the mapped time slice position at 1-second intervals. The average heat intensity of the infrared thermal image is calculated according to the 8×8 grid region and mapped to the standard spatial grid. The energy consumption data is summed at the 1-second time slice boundary to obtain the power consumption value for each slice. Based on this, the three types of data are merged through indexing to form a composite physical quantity matrix (number of time slices × number of spatial grids). Each row of the matrix contains the average temperature, thermal image intensity distribution, and energy consumption value for the corresponding time slice. Based on the complete consistency between this matrix and the cooking parameter set in the spatiotemporal dimensions, the accuracy of residual calculation and the stability of deviation identification can be significantly improved.

[0106] S4.3: Perform point-by-point difference operation on the spatiotemporally aligned actual operating data set and the cooking parameter set, and use the sliding window variance statistics method to calculate the residual fluctuation amplitude in each time slice to generate an original deviation feature vector containing temperature deviation value, heat distribution unevenness and energy consumption deviation rate.

[0107] The spatiotemporally aligned actual operational data set and cooking parameter set are processed using point-by-point difference operations under a unified time slice and spatial grid to calculate the instantaneous residual values ​​of the corresponding physical quantities. A sliding window is constructed for the residual sequence with a preset time slice width. Within each window, variance statistics are used to extract the residual fluctuation amplitude. The variance calculation formula is as follows: Multiply by the sum of squared residuals, where the sum of squared residuals is The square of, For the first The difference between actual and predicted values ​​for each time slice is analyzed. The variance results are mapped to three feature dimensions: temperature deviation, thermal unevenness, and energy consumption deviation rate. An original deviation feature vector is constructed using a multidimensional residual matrix. For the temperature deviation dimension, the residual variance of the cavity temperature curve is extracted and converted to thermodynamic units. For the thermal unevenness dimension, the variance of the local temperature difference is calculated by combining the residual of the infrared thermal image matrix. For the energy consumption deviation rate dimension, an energy efficiency deviation index is formed by combining the residual variance of the energy consumption trajectory with the ratio of the baseline energy consumption. Through residual difference quantification and multidimensional feature combination processing, the results of the previous step are transformed into an original deviation feature vector containing temperature deviation, thermal unevenness, and energy consumption deviation rate, thus preparing input data for identifying significant deviation periods and parameter dimensions.

[0108] S4.4: Based on the original deviation feature vector, an adaptive threshold segmentation algorithm is applied to identify continuous time periods that exceed the allowable tolerance range, and gradient direction analysis is combined to determine the key control parameter dimensions that cause the deviation, so as to generate the deviation identification result.

[0109] Based on the temperature deviation, heat distribution unevenness, and energy consumption deviation components in the original deviation feature vector, the adaptive threshold segmentation algorithm module is invoked to obtain the allowable tolerance benchmark value corresponding to each component, and a dynamic threshold matrix is ​​formed by combining it with the real-time operating condition correction coefficient. Element-by-element comparison operations are performed between the multidimensional components of the original deviation feature vector within each time slice and the dynamic threshold matrix to generate a Boolean mask matrix indicating points exceeding the allowable tolerance range. Continuous segment identification processing is performed on the Boolean mask matrix, and isolated out-of-tolerance points are eliminated using the minimum duration constraint, forming a set of continuous out-of-tolerance time periods. Gradient direction analysis is performed on the deviation curves within the set of continuous out-of-tolerance time periods, and the correlation between the direction vector of deviation change and the control parameter change vector of the cooking parameter set is calculated to determine the key control parameter dimension causing the deviation. An index mapping relationship is established between the set of continuous out-of-tolerance time periods and the corresponding key control parameter dimensions, and the deviation identification result is output. Through the combined processing of adaptive threshold segmentation and gradient direction analysis, the residual fluctuation amplitude data from the previous step is transformed into significant deviation periods and key parameter dimension information that can be used to locate the error source, achieving the expected technical effect of rapid and accurate identification of the model deviation source.

[0110] For example, in a certain operation of a commercial multi-energy oven, the original deviation feature vector shows temperature deviation values ​​ranging from 2.4℃ to 7.9℃, heat distribution unevenness ranging from 0.12 to 0.35, and energy consumption deviation rate ranging from 0.08 to 0.29. Allowable tolerance baseline values ​​are set to 3℃, 0.15, and 0.1 respectively, and a threshold coefficient of 1.1 is corrected when the energy fingerprint fluctuation intensity index is 1.25, forming a dynamic threshold matrix. After performing comparison operations on each time slice, a Boolean mask matrix is ​​obtained. After processing with a minimum duration constraint (greater than 5 seconds), two periods of continuous out-of-tolerance are identified. Gradient direction analysis is performed on the temperature deviation curves during these two periods, revealing a maximum correlation coefficient of 0.87 with the change vector of the heating power setpoint, and correlation coefficients of 0.43 and 0.31 with the damper opening adjustment and time segment threshold, respectively. The heating power setpoint is identified as a key control parameter dimension. The significant deviation periods output in this step are from 420 to 465 seconds and from 780 to 810 seconds. The key control parameter dimension is the heating power setpoint. In the subsequent continuous error determination in S5, this information is directly used to determine whether to trigger local model fine-tuning, thereby significantly improving the prediction accuracy.

[0111] Step S5: Based on the deviation identification result, determine whether the prediction error under the same type of energy fingerprint input for a consecutive preset number of times exceeds a set threshold. If it does, trigger a fine-tuning command. Specifically, this includes: S5.1: Obtain the single prediction error data in the deviation identification results, and use the sliding time window algorithm to perform time-series arrangement processing on the single prediction error data corresponding to the same type of energy fingerprint input for a consecutive preset number of times, so as to generate a continuous error data queue containing historical error sequences.

[0112] Based on the deviation identification results output by the preceding S4.4, the data parsing module is called to extract the numerical components of the single prediction error data, and the data is classified and stored according to the energy fingerprint category to form a grouped data table.

[0113] For each energy fingerprint in the grouped data table, a sliding time window with a fixed width is established. The window width parameter is determined by the product of the control period and the sampling frequency. Multiple prediction error data under the same type of energy fingerprint input conditions are arranged sequentially on the time axis.

[0114] A window progression operation is performed on the error data entries in the sliding time window to generate a time series index. The corresponding prediction error value and parameter dimension identifier are bound to each time slice to ensure the traceability of subsequent series processing.

[0115] The total data length of the continuous error sequence is calculated using the sequence index and the prediction error value. When the data length is less than the preset consecutive number threshold, the missing time slice data is automatically filled with zero values ​​to avoid statistical bias.

[0116] The amended time series are reassembled in time slice order to form a continuous error data queue containing historical error sequences, providing a unified input data structure for subsequent noise suppression and trend extraction steps. Through sliding time window arrangement and amending, the deviation identification results from the previous step are transformed into a continuous error data queue that meets the temporal consistency requirements, achieving a structured caching effect for historical prediction errors under multi-energy fingerprint conditions.

[0117] For example, in a multi-energy oven control system, the numerical characteristics of the deviation identification result output include a temperature deviation of 2.8°C, a heat distribution non-uniformity of 0.35, and an energy consumption deviation rate of 0.12. The system detects that the energy fingerprint category is gas mode, sets the sliding time window width to 5 control cycles, the sampling frequency to 1Hz, and calculates the number of samples corresponding to the window width to be 5. The 5 single prediction error values ​​are arranged in time series as [2.8, 2.5, 3.0, 2.7, 2.9], and bound to the corresponding parameter dimension identifier "heating rate". After performing the window progression operation, a time series index [1, 2, 3, 4, 5] is generated, and the corresponding error values ​​are stored in pairs with the indexes. When the historical consecutive number threshold is detected to be 7 and the current sequence length is 5, the error value of the two missing time slices is filled to 0.0, resulting in a complete continuous error data queue [2.8, 2.5, 3.0, 2.7, 2.9, 0.0, 0.0]. In this embodiment, the continuous error data queue fully preserves the structured information of the prediction error changing over time under the gas mode, which can be used subsequently for weighted moving average filtering to calculate the cumulative error feature vector.

[0118] S5.2: Based on the historical error sequence in the continuous error data queue, the weighted moving average filtering algorithm is used to suppress noise and extract trends from the historical error sequence, generating a cumulative error feature vector.

[0119] For the historical error sequence in the continuous error data queue generated by step S5.1, a vector of single-prediction residual values ​​containing multiple time slices is received as input. These are sequentially loaded into the filtering unit according to their time slice indices. The core module of the weighted moving average filtering algorithm is invoked, and corresponding weight values ​​are applied to each historical residual value based on the stability coefficient of the energy fingerprint type to which the time slice belongs, forming a weight matrix with energy fingerprint adaptability correction. After multiplying the corresponding elements of the weight matrix and the historical residual value vector, an accumulation operation is performed to obtain the weighted residual sum. The denominator weight normalization module is then invoked to divide the weighted residual sum by the sum of the weight coefficients to complete the average value calculation, as shown in the following formula: in It is a weighted moving average. For the first Energy stability weighting coefficients corresponding to each time slice Predict the residual value for this time slice. Input the obtained weighted average result as the trend baseline value into the trend extraction module, and calculate the trend rate of change over continuous time slices using the first-order difference operator, as shown in the following formula: in For the current time slice index, This is a weighted moving average of adjacent time slices. Based on the sign and amplitude of the rate of change, the noise suppression threshold module is invoked to remove low-amplitude high-frequency oscillation components, retaining low-frequency components with stable trends, and generating a cumulative error feature vector. Through the above weighted moving average and trend extraction processing methods, the continuous historical residual sequence from the previous step is transformed into quantitative features characterizing the model's predictive stability under the current energy fingerprint, thus providing accurate input for the dynamic threshold calculation in S5.3.

[0120] For example, in an embedded oven supporting dual-energy modes, the continuous error data queue contains the predicted residual values ​​for 20 time slices, in degrees Celsius. The set of weight coefficients corresponding to the energy fingerprint is [0.9, 0.85, 0.88, 0.92]. Multiplying the residual values ​​of the 20 time slices by the corresponding elements of the weight coefficients and summing them yields a weighted residual sum of 18.4. Dividing this by the total sum of the weight coefficients (17.9) yields a weighted average of 1.0279℃, which is used as the trend baseline value input to the trend extraction module. First-order differencing is performed on the baseline values ​​of adjacent time slices. If the difference is less than 0.05℃ and persists for 3 time slices, it is considered noise and suppressed. If the difference is greater than 0.1℃ and persists for 5 time slices, it is considered a trend change and retained. The resulting cumulative error feature vector exhibits stable fluctuations under the same energy fingerprint, with the trend feature vector amplitude between 1.02℃ and 1.15℃, significantly improving the robustness and accuracy of subsequent dynamic threshold determination.

[0121] S5.3: Obtain the error components of each dimension in the cumulative error feature vector, and use the dynamic threshold calculation strategy to adaptively correct the preset basic threshold according to the heat sensitivity coefficient of the current recipe identifier and the fluctuation intensity index of the energy fingerprint, so as to generate a real-time judgment threshold boundary for the current working condition.

[0122] The system receives a multi-dimensional data structure of the cumulative error feature vector and uses the error components of each dimension as input parameters for the current operating condition.

[0123] The heat sensitivity coefficient corresponding to the recipe identifier is parsed, and the coefficient is read from the recipe attribute database to quantify the sensitivity of the cooking target to temperature fluctuations.

[0124] The fluctuation intensity index of the energy fingerprint is extracted and obtained from the energy physical characteristic analysis module to characterize the disturbance intensity of energy input changes on model predictions.

[0125] The dynamic threshold calculation strategy is invoked, and the preset base threshold is multiplied and corrected by the heat sensitivity coefficient. Then, the corrected result is weighted and superimposed with the fluctuation intensity index to form the operating condition adjustment value, as shown in the following formula: in, To determine the threshold boundary in real time, To preset the basic threshold, The thermal sensitivity coefficient, These are the weighting coefficients. This is the volatility intensity index.

[0126] Based on the operating condition adjustment value, the error components of each dimension are expanded or contracted to ensure that the threshold boundary is more stringent in the sensitive parameter dimension and more lenient in the stable parameter dimension.

[0127] The processed multidimensional threshold set is encapsulated into a real-time threshold boundary determination data structure, which serves as the direct input for subsequent error exceeding the limit judgment.

[0128] By using the aforementioned dynamic threshold correction method, the cumulative error feature vector from the previous step is transformed into a real-time judgment threshold boundary for the current operating condition, thereby improving the accuracy of adaptive error judgment under multi-energy conditions.

[0129] For example, in the gas mode of a commercial oven, the preset basic threshold is configured as 0.15, the heat sensitivity coefficient corresponding to the recipe identifier is 1.2, the energy fingerprint fluctuation intensity index is 0.08, and the weighting coefficient is set to 0.5. The real-time judgment threshold boundary calculated according to the formula is 0.206. This threshold boundary is applied to the error components of each dimension of the cumulative error feature vector. Specifically, the temperature deviation dimension triggers an error exceeding the limit judgment when it reaches 0.21 in the continuous sampling window, the heat distribution unevenness dimension remains at 0.19 without triggering, and the energy consumption deviation rate dimension triggers a slight correction at the critical value of 0.205. Verification shows that the dynamic threshold calculation in this scenario significantly improves the detection accuracy of sensitive parameters while maintaining tolerance for stable parameters, thereby optimizing the local model fine-tuning trigger mechanism and improving the system's adaptability and stability in gas mode.

[0130] S5.4: Based on the cumulative error feature vector and the real-time judgment threshold boundary, perform a multi-dimensional vector comparison operation to compare the magnitude of the cumulative error feature vector with the real-time judgment threshold boundary to generate an error over-limit Boolean flag indicating whether the prediction accuracy has failed.

[0131] In the multi-dimensional vector comparison operation based on the cumulative error feature vector and the real-time judgment threshold boundary, the real-time judgment threshold boundary output in the previous step S5.3 and the cumulative error feature vector obtained in S5.2 are aligned and mapped according to the corresponding parameter dimensions to ensure that each dimension error component corresponds one-to-one with its judgment boundary.

[0132] The magnitude of the aligned cumulative error eigenvector is calculated by performing a sum of squares on all dimensional error components using the Euclidean norm formula and taking the square root to generate the magnitude of the cumulative error vector. This calculation process can be expressed as follows: in The first eigenvector representing the cumulative error eigenvector Each dimension component This indicates the time slice index for the current operating condition. This represents the magnitude of the cumulative error vector.

[0133] The above-mentioned modulus calculation result is compared with the overall amplitude of the real-time threshold boundary using a vector scale. The multidimensional vector comparison operator is called to determine whether the modulus exceeds the threshold amplitude and a preliminary over-limit Boolean flag is generated.

[0134] To prevent misjudgments caused by instantaneous spikes, the Boolean flag of exceeding the limit is logically smoothed. The current judgment result is compared with the judgment results of the last three time slices by a majority logical vote. If the continuous exceeding ratio reaches the preset condition, the true value is maintained in the smoothed result.

[0135] The smoothed Boolean flag is used as the final output for determining the failure of prediction accuracy and is passed to the subsequent local model fine-tuning trigger logic.

[0136] By using multidimensional vector magnitude calculation and real-time threshold comparison, the cumulative error feature vector from the previous step is transformed into a Boolean flag that can be directly used to determine model accuracy failure, thus achieving accurate identification of error exceeding limits.

[0137] For example, in an embedded oven that supports dual energy switching between electricity and gas, the four dimensions of the cumulative error feature vector correspond to temperature deviation, heat distribution unevenness, energy consumption deviation rate, and time-period stability, with values ​​of 1.8, 2.3, 0.9, and 1.5, respectively. The real-time threshold boundary amplitude calculation result is 2.0. Applying the Euclidean norm formula to the error components of each dimension, the calculated modulus is 3.29, exceeding the threshold of 2.0, generating a preliminary out-of-limit Boolean truth flag. After a majority logic vote in three consecutive time slices where all values ​​are true, the prediction accuracy is finally determined to be faulty, triggering a subsequent local model fine-tuning process. In this embodiment, through the above determination process, the system successfully identifies the cooking curve prediction failure state during hybrid energy switching, significantly improving model adaptability, enhancing the stability of the cooking temperature field, and optimizing energy efficiency.

[0138] S5.5: In response to the true value of the error exceeding limit Boolean flag, the local optimization process is activated when the error exceeding limit Boolean flag is true using state machine transition logic, so as to generate the fine-tuning instruction that drives the main track model to update the weights.

[0139] Based on the true state input of the error exceeding the Boolean flag, the state machine mapping table of the local optimization process is invoked to parse the current model running stage and the corresponding execution path. The true value of the Boolean flag is compared with the trigger condition field in the state machine mapping table to generate a trigger event descriptor containing the current energy fingerprint category and recipe identifier. Event priority sorting is performed on the trigger event descriptor, and the optimization process branch node with the highest priority is selected within the state machine based on the sorting result. Using the selected optimization process branch node, a set of model weight update strategies related to the significant deviation period and key parameter dimensions is retrieved, and the update strategy matching the main track model weight structure is extracted. The weight update step size, freeze mask mode, and gradient calculation configuration items involved in the update strategy are encapsulated into a core control parameter set for fine-tuning instructions. By matching and encoding the above core control parameter set with the trigger event descriptor, fine-tuning instructions that can be directly executed on the main track model for weight iteration are generated.

[0140] By comparing the state machine transition logic with the triggering conditions, the error exceeding the limit judgment result of the previous step is transformed into a fine-tuning instruction with real-time execution capability, realizing rapid and targeted weight adjustment of the main track model under multi-energy fluctuation conditions.

[0141] For example, in the operating scenario of a commercial multi-energy oven, the cumulative error feature vector magnitude of three consecutive batches of gas energy fingerprint inputs is 5.8, while the real-time judgment threshold boundary for this condition is 4.5, satisfying the error exceeding the limit condition. The state machine mapping table maps this truth flag to the highest priority "Gas-High Thermosensitive Recipe" fine-tuning process branch. The update strategy bound to this branch node is to freeze a subset of neurons related to the heating rate setpoint and perform single-step gradient updates with a learning rate of 0.002. The core control parameter set of the fine-tuning instruction includes a 1024-bit frozen mask bitmap, a gradient calculation batch size of 64, and a weight update step size of 0.002. The above parameter set is merged and encoded with the event descriptor "Gas Energy Fingerprint + Recipe ID245" to generate a fine-tuning instruction that can directly trigger weight iteration in the current main track model instance. After executing this instruction, the oven's heating rate prediction accuracy under the same energy input conditions is significantly improved, the cavity temperature field deviation is close to the target cooking curve, and the energy consumption change tends to be stable.

[0142] Step S6: In response to the fine-tuning instruction, freeze the subset of neurons in the main trajectory model that are strongly correlated with specific parameters, and perform single-step gradient updates using the current batch running data to generate incremental patches. Specifically, this includes: S6.1: Based on the deviation identification results, locate the subset of neurons strongly correlated with specific cooking parameters whose prediction errors exceed a set threshold from the fully connected layer topology of the main track model, and generate a gradient freeze mask for the subset of neurons to lock the weight states of unrelated neurons to prevent catastrophic forgetting.

[0143] Based on the deviation identification results, the fine-tuning instruction triggered when the Boolean flag indicating that the error exceeds the limit is true is set, and the input object is set as the association information of the fully connected layer topology of the master track model and the specific cooking parameters whose prediction error exceeds the set threshold.

[0144] The parameter association mapping table is invoked to calculate the correlation coefficient between the specific cooking parameter components in the recognition results and the weight matrices of each layer of the main track model, forming a parameter-neuron association matrix.

[0145] Based on this correlation matrix, a saliency ranking algorithm is used to extract the set of neuron indices with correlation coefficients greater than a static threshold, which are then identified as a subset of neurons strongly correlated with the target cooking parameters.

[0146] A position index mapping is performed on the located subset of neurons to construct a gradient freeze mask matrix. Structurally, this matrix selectively enables gradient propagation paths by filling the positions of unrelated neurons with zero values ​​and the positions of related neurons with one value.

[0147] By combining the gradient freezing mask matrix with the current weight state of the main track model, a protective masking mechanism is generated to lock the weight states of irrelevant neurons, so as to avoid catastrophic forgetting in the subsequent weight update process.

[0148] By using a gradient freezing mask to lock the localization results of the previous step, it ensures that the forward propagation calculation in S6.2 is performed only on the restricted paths of the relevant neurons, thus achieving the precise target range required for local weight correction.

[0149] For example, in the operating environment of a commercial multi-energy oven, the cooking mode corresponding to the gas main heating was detected. The prediction error of its heating power setting value was 2.4, exceeding the set threshold of 1.8. In the correlation matrix calculation, the correlation coefficients between this parameter and some neurons in the 2nd and 4th layers of the fully connected layer were 0.82 and 0.87, respectively. The static threshold was set to 0.8, thus selecting 12 neurons in the 2nd layer and 8 neurons in the 4th layer as a subset of strongly correlated neurons. When constructing the gradient freezing mask matrix, in the mask of size 256×1, the index position of the corresponding strongly correlated neuron was assigned a value of 1, and the remaining positions were assigned a value of 0. This mask is used to mask irrelevant weights in subsequent single-step gradient updates, so that the update process is limited to the weight values ​​of these 20 neurons. The gradient freezing mask generation formula is defined by the following explicit mathematical expression: in For the code vector Each element has a value of 1 (allowing gradient updates) or 0 (freezing weights). Represents the correlation coefficient function. For neuron indexing, The static correlation threshold, This is the frozen mask matrix. After applying this mask, the model updates the heating power setpoint only on key relevant weights in the gas-fired main heating scenario, avoiding abrupt changes in the weights of other parameters such as damper opening, and achieving stable adaptive correction under the background of energy input fluctuations.

[0150] S6.2: Utilize the gradient freeze mask to perform forward propagation calculation on the feature vector of the cavity temperature field evolution curve in the current batch of running data, and obtain the local intermediate layer output tensor containing only the activation state of the subset of neurons to be updated, providing a limited computational graph path for subsequent backpropagation.

[0151] Using the gradient freeze mask generated in the preceding step S6.1 as a limiting condition for the weight update range, a restricted forward propagation calculation is performed on the feature vector of the cavity temperature field evolution curve in the current batch of running data.

[0152] When the feature vector of the cavity temperature field evolution curve is input into the neurons of each layer of the main track model, a gradient freezing mask is applied to close the activation paths of non-target neurons in the computation graph, so that only neurons strongly correlated with specific cooking parameters keep their computation channels open.

[0153] During the tensor flow between the convolutional layer and the fully connected layer, a freeze mask is used to nullify the weight product of unrelated neurons, ensuring that the output tensor does not contain the contribution of non-target neurons.

[0154] When calculating the feature map of the local intermediate layer, the freeze mask is applied to the corresponding index position of the multidimensional tensor, and the output data within the scope of the application is extracted into the local intermediate layer output tensor containing only the activation state of the subset of neurons to be updated.

[0155] The output tensors of the local intermediate layers are structured and labeled with timestamps and energy fingerprint identifiers so that the backpropagation in the subsequent step S6.3 can extract the local gradient matrix based on the restricted computation graph path.

[0156] By combining gradient freezing masking with restricted forward propagation, the neuron localization results from the previous step are transformed into local intermediate layer output tensors that can be used for local backpropagation, thereby achieving accurate construction of the local computation graph and strict limitation of the gradient update region.

[0157] S6.3: Based on the residual loss function between the output tensor of the local intermediate layer and the cavity temperature field curve label, execute a single-step backpropagation algorithm to calculate the local gradient matrix of the subset of neurons to be updated, thereby quantifying the correction direction and magnitude of the model parameters under the current energy fingerprint input.

[0158] Based on the residual loss function between the output tensor of the local intermediate layer and the label of the cavity temperature field curve, a computational graph path with a freeze mask constraint is loaded, and the residual matrix is ​​obtained by performing element-wise difference operations on the output tensor of the local intermediate layer and the label data.

[0159] The loss calculation logic is invoked to perform a square operation on the residual matrix and then sum them to form the local mean squared error loss value. The formula used is as follows: in Indicates the number of valid samples. This represents the predicted value output by the local intermediate layer. This represents the actual label value of the corresponding sample.

[0160] A single-step backpropagation algorithm is performed on the loss value. Based on the chain rule, the loss gradient is propagated to the previous layer along the open path of the frozen mask, while the gradient matrix of the subset of neurons to be updated is calculated.

[0161] The direction correction coefficient matrix is ​​obtained by multiplying each element in the gradient matrix with the feature weight corresponding to the energy fingerprint input.

[0162] Perform sign operations and amplitude normalization on the direction correction coefficient matrix to extract the correction direction (positive increase or negative decrease) and correction amplitude (normalized to a preset range) of the weights of each neuron, forming a quantized local gradient matrix.

[0163] Through the above processing method, the local intermediate layer output tensor and label data of the previous step are transformed into direction and magnitude indicators that can be directly used for weight updates, thereby realizing the directional correction of the main track model parameters under the current energy fingerprint.

[0164] For example, in a commercial kitchen's multi-energy mode, the main gas heating power is 3000W, and the electric auxiliary heat preservation power is 800W. At this time, the energy fingerprint vector is a fixed-length 128-dimensional vector, the local intermediate layer output tensor size is (64×128), the cavity temperature field curve sampling frequency is 1Hz, and the sample size n is 64. The predicted value matrix and the label value matrix are differenced to obtain the residual matrix. The mean square error L after summing the squared residuals is calculated to be L≈0.071. After performing a single-step backpropagation, the gradient matrix size is (64×128). Multiplying it by the feature weights corresponding to the energy fingerprint vector yields the direction correction coefficient matrix, which is normalized to the interval [-0.005, 0.005], where positive values ​​correspond to increased weights and negative values ​​correspond to decreased weights. Ultimately, in this energy mode, the model's matching accuracy for the heating power setting and time segment threshold is significantly improved, the cavity temperature fluctuation is controlled within ±0.3℃, the energy consumption deviation rate is significantly reduced, and the output stability of the cooking process is greatly improved.

[0165] S6.4: Perform a weight iterative update operation based on the local gradient matrix and the preset adaptive learning rate coefficient to generate a temporary weight matrix containing the latest adaptation information, and perform element-wise subtraction operation between the temporary weight matrix and the original baseline weight matrix before freezing to extract the weight difference data block in pure incremental form.

[0166] The local gradient matrix and preset adaptive learning rate coefficients generated in the previous steps are obtained to construct an update operator matrix for weight iteration. Based on this update operator matrix, element-wise operations are performed on the temporary weights. The weight values ​​are adjusted positively or negatively by adjusting the gradient direction and magnitude, so that the adjusted weight matrix can more closely approximate the observed data of the cavity temperature field curve under specific energy fingerprint input conditions. During this iteration, the original baseline weight matrix before freezing is used as a comparison benchmark. The matrix difference operation module is called to perform subtraction operations one by one at the same neuron index position.

[0167] in, For weighted difference data blocks, This is a temporary weight matrix. This is the original baseline weight matrix. The difference matrix is ​​limited to contain only the row and column units corresponding to the subset of neurons being updated to ensure the sparsity and specificity of the difference data. The difference matrix extraction module outputs weight difference data blocks in pure incremental form, which serve as the raw material for subsequent incremental patch generation. Through this chain-iteration and difference processing method, the local gradient matrix from the previous step is transformed into storable and indexable weight difference data, achieving the desired technical effect of minimizing storage load and fast loading for weight updates.

[0168] For example, in a multi-energy input scenario, the maximum magnitude of the local gradient matrix is ​​0.012, the preset adaptive learning rate coefficient is set to 0.005, and the absolute value of each element of the update operator matrix after element-wise multiplication is less than 6 × 10⁻⁶. 5 The update operator matrix is ​​then element-wise added to the original baseline weight matrix to obtain a temporary weight matrix. The temporary weight matrix and the original baseline weight matrix are then calculated using the aforementioned difference formula at their index positions, resulting in a weight difference data block with 128 non-zero elements, a very small proportion of the total elements, verifying the sparsity preservation effect. When similar energy fingerprints reappear, this weight difference data block is quickly superimposed onto the baseline weight matrix of the main track model. The deviation between the output cooking power curve and the target temperature field curve is significantly reduced during the heating stage, the heat distribution uniformity is greatly improved, and the energy utilization rate is significantly increased.

[0169] S6.5: Perform sparse encoding and quantization compression on the weight difference data block to generate an incremental patch file with hash-like index characteristics. This patch file will be used as the basis for adaptive correction when the same energy fingerprint appears in the future, and will be automatically loaded and superimposed on the main track output, thus completing the local model fine-tuning loop.

[0170] Step S7: Establish an index mapping relationship between fingerprints and patches, and store the incremental patch in the calibration cache so that it can be automatically loaded and superimposed on the main rail output when the same energy fingerprint appears later, forming an adaptively corrected combination of cooking parameters. Specifically, this includes: S7.1: Obtain the incremental patch file with hash-like index characteristics generated in the previous steps and the corresponding fixed-length energy fingerprint vector, and perform hash key-value mapping processing on the fixed-length energy fingerprint vector based on the doubly linked list data structure to generate fingerprint index entries containing unique address pointers.

[0171] Based on the incremental patch file with hash-like index characteristics generated in the previous steps and the corresponding fixed-length energy fingerprint vector, the memory data receiving interface is called to obtain the complete patch file byte stream and the numerical sequence of the fixed-length energy fingerprint vector. During the receiving process, a cyclic redundancy check algorithm is used to verify the data packets in real time to ensure data integrity.

[0172] After obtaining the numerical sequence of the fixed-length energy fingerprint vector, the hash function calculation module is called to calculate the hash value of the fingerprint vector. A hash algorithm with high distribution uniformity is used to map the vector to a fixed-length hash key value so that it can be quickly searched in the subsequent index structure.

[0173] After the hash value is calculated, a linked list data structure containing head and tail pointers is created using doubly linked list initialization logic, and an independent memory storage unit is allocated for each linked list node to ensure that the insertion and deletion of index entries can be completed in constant time.

[0174] During the construction of the linked list node, the hash key value of the fixed-length energy fingerprint is used as the primary key field of the node, and a unique address pointer is generated. This pointer points to the reserved storage location of the incremental patch file in the calibration cache, providing direct addressing capability for subsequent writing of patch data units.

[0175] During the binding process between the primary key field of the linked list node and the unique address pointer, the index mapping update function is called to insert the newly generated fingerprint index entry into the corresponding position of the doubly linked list, and the node order is adjusted according to the current linked list node sorting strategy to ensure that the hash key value can match the corresponding patch file address under the shortest path during the search process, thereby achieving efficient retrieval.

[0176] By using the doubly linked list hash key-value mapping method described above, the incremental patch file generated in the previous step is bound to the fixed-length energy fingerprint vector as a fingerprint index entry containing a unique address pointer, thus achieving the technical effect of fast local addressing and efficient management.

[0177] For example, in the parameter mapping system of a commercial multi-energy oven, the fixed-length energy fingerprint vector is represented by 128-bit binary encoding. After processing by a hash function, a 16-byte hash key value is generated, such as 0x3FA5C1E4B8972D6C. This hash key value is stored as the primary key field of a linked list node. The system assigns a unique address pointer to this node in the doubly linked list, such as 0x7FFDE120. This address pointer points to a 256KB data block reserved in the calibration cache. After the linked list insertion operation is completed, the average time to retrieve the patch file using this hash key value is significantly reduced, with the retrieval latency decreasing from 1.5 milliseconds to 0.3 milliseconds. The cache hit rate is thus significantly improved, demonstrating the performance advantage of fast online energy fingerprint matching. In this scenario, when the same energy fingerprint reappears, the system can immediately hit the linked list and load the corresponding incremental patch file, achieving efficient master rail model correction and cooking parameter mapping optimization.

[0178] S7.2: Receive the fingerprint index entry and the incremental patch file, divide the incremental patch file into fixed-size data blocks using a non-volatile memory paging write protocol, and attach a cyclic redundancy check code to each data block to generate a standardized patch data unit with an integrity protection mechanism.

[0179] S7.3: Based on the standardized patch data unit and the fingerprint index entry, perform a memory address binding operation to point the unique address pointer to the physical storage location of the standardized patch data unit in the calibration cache, so as to construct a direct addressing mapping table from fingerprint to storage address.

[0180] Based on the standardized patch data unit and the fingerprint index entry, the physical address information of the standardized patch data unit in non-volatile memory is read by calling the memory address mapping interface, and this information is matched one-to-one with the unique address pointer in the fingerprint index entry. Using address binding control logic, the value of the unique address pointer is rewritten to the actual physical address value, ensuring that the pointer can directly locate the storage location of the patch data unit during cache retrieval. Through an address mapping optimization mechanism, the physical address is aligned during the binding process, correcting any misaligned address offsets to the storage page boundary to improve cache access efficiency. A direct addressing mapping structure is adopted, storing the unique address pointers of all fingerprint index entries and their corresponding physical address records in a high-speed mapping table, constructing a direct addressing relationship from energy fingerprints to standardized patch data units. Through the binding of unique address pointers and physical addresses and the construction of the mapping table, the standardized patch data unit generated in the previous step is transformed into a data location structure that can be retrieved instantly, achieving an efficient and accurate mapping effect between energy fingerprints and patch data in the calibration cache.

[0181] For example, in the embedded controller of a commercial oven, the fixed-length energy fingerprint vector is set to 64 bytes. The incremental patch file is divided into standardized patch data units of 256 bytes each after sparsification and compression. The page size of the non-volatile memory is 1024 bytes, and each page can hold four data units. The physical address of the patch data unit in memory is read through the memory address mapping interface. For example, the initial value of the unique address pointer of a fingerprint index entry is logical offset 500, which is adjusted to physical address 1024 after page alignment operation. The page alignment formula is as follows: in, For the aligned physical address, For page size, This is a logical offset. The logical offset 500, after page alignment, yields a physical address of 1024, which is then bound to a unique address pointer. This creates a direct mapping from the 64-byte energy fingerprint vector to the 1024 physical address in the mapping table. Testing showed that in scenarios where the energy fingerprint retrieval hits a patch, the time to retrieve the patch data unit was significantly reduced, the system's cooking parameter loading latency was significantly reduced under multi-energy switching conditions, and the thermal stability was significantly improved.

[0182] S7.4: Call the direct addressing mapping table and dynamically manage the remaining space of the calibration cache based on the least recently used replacement strategy. When the cache capacity is detected to reach a preset threshold, the old standardized patch data units that are accessed infrequently are automatically replaced to generate an updated calibration cache with optimal space utilization.

[0183] The unique set of address pointers in the direct addressing mapping table is used as the input for cache space management. The remaining available space value of the calibration cache is sequentially read to match the storage capacity of the current standardized patch data unit. Based on the remaining available space value, cache space occupancy is calculated using the following formula: in, For cache utilization, This represents the amount of cache space already used. The total cache capacity is defined. The occupancy rate is compared vector-wise with a preset threshold set to generate an over-occupancy detection result. The least recently used replacement strategy is invoked to statistically process the access frequency of index entries in the direct addressing mapping table, and the standardized patch data units with low-frequency access are located based on the access frequency sorting result. Logical deletion operations are performed on the low-frequency access data units, releasing their corresponding physical storage space, and the direct addressing mapping table is updated to remove the corresponding unique address pointer entry. The released space value is compared with the space required by the current patch data unit; if insufficient, the low-frequency access eviction process is recursively executed until the storage requirements of the patch data unit are met. Through access frequency statistics and space release processing, the result of the previous step is transformed into an updated calibrated cache with optimal space utilization, achieving dynamic and sustainable management of cache resources.

[0184] For example, in an embedded multi-energy oven control system, the incremental patch data unit corresponding to the fixed-length energy fingerprint vector has a capacity of 4MB, the total calibration cache capacity is 32MB, the currently used capacity is 29MB, and the preset occupancy threshold is 0.85. The occupancy calculation result is 0.906, indicating an over-limit status. The least recently used replacement strategy is invoked to sort the 20 index entries in the direct addressing mapping table by access frequency. Patch data units with access frequencies less than 5 times are selected, and their occupied space is released sequentially. Each released data unit updates the mapping table and removes the corresponding unique address pointer. In this example, by releasing two sets of low-frequency access patch data units totaling 5MB, the remaining space is 8MB, which is greater than the 4MB space required by the current patch data unit. This significantly improves the utilization rate of the calibration cache space after the update and allows for rapid patch loading during subsequent energy fingerprint matching, ensuring the stability and accuracy of cooking parameter output.

[0185] S7.5: Based on the standardized patch data units stored in the updated calibration cache, when a new fixed-length energy fingerprint vector input is detected, a fast retrieval and reading operation is performed, and the read weight difference data block is superimposed on the baseline weight matrix of the main track model in real time to output the final cooking parameter combination after adaptive correction.

[0186] Step S8: For multi-energy mixed input scenarios, the energy fingerprints of multiple independent energy sources are weighted and fused according to real-time power ratios to generate a composite fingerprint. The final control command is then output using the master track model combined with thermal hysteresis compensation and power switching jitter suppression algorithms. Specifically, this includes: S8.1: Obtain the real-time power value of each independent energy source and the corresponding independent energy source fingerprint vector at the current moment, and calculate the power weight coefficient of each independent energy source in the total input power based on the real-time power value to generate a set of power weight coefficients for characterizing the contribution of multiple energy sources.

[0187] S8.2: Perform a weighted summation operation on the independent energy fingerprint vector and the power weight coefficient set to linearly superimpose the multidimensional physical layer features according to the energy contribution ratio, thereby generating a composite fingerprint vector characterizing the transient characteristics of the hybrid energy.

[0188] S8.3: Based on the composite fingerprint vector, retrieve the fingerprint and patch index mapping relationship in the calibration cache. If a match is found, load the corresponding incremental patch data and superimpose the incremental patch data onto the pre-stored baseline weight parameters of the master track model to generate a dynamically calibrated real-time master track model.

[0189] S8.4: The composite fingerprint vector and the recipe identifier selected by the user are jointly input into the real-time master track model for forward inference calculation to calculate the initial multi-source control parameter set including the basic heating curve, thermal hysteresis compensation and power switching jitter suppression.

[0190] It should be noted that the thermal hysteresis compensation algorithm is based on the deviation between the actual furnace temperature response curve and the target curve. It uses a proportional-integral-derivative (PID) controller with feedforward to calculate the compensation amount. The proportional term is used to eliminate steady-state temperature deviation, the integral term is used to suppress long-term thermal accumulation hysteresis, and the derivative term is used to reduce temperature overshoot. The power switching jitter suppression algorithm uses a first-order low-pass filter and a rate limiter in series to post-process the power command output by the master rail model. The time constant of the first-order low-pass filter is dynamically adjusted according to the power fluctuation intensity index in the current composite fingerprint. The rate limiter constrains the power change rate of adjacent control cycles within a preset safe slope, thereby avoiding drastic fluctuations in the cavity temperature field caused by power jumps during multi-energy switching.

[0191] The composite fingerprint vector and the user-selected recipe identifier are processed by a multimodal feature alignment module to generate a joint input tensor containing transient energy physical features and the thermodynamic requirements of the dish. This joint input tensor is then processed by the input preprocessing layer of the real-time mainline model to normalize the tensor and adjust the feature domain weights, ensuring the matching scale of features from different sources in the model computation. The normalized result is fed into the deep feature extraction network of the mainline model, passing through convolutional layers and recursive units to extract a high-dimensional hidden state graph reflecting the difference between the dynamics of energy supply and the expected cooking temperature curve. This high-dimensional hidden state graph is input into the multi-branch decoder of the mapping sub-model. The basic heating curve branch outputs the heating power-time function for the current energy combination, the thermal hysteresis compensation branch calculates the adjustment amount for furnace temperature overshoot and delay, and the power switching jitter suppression branch generates suppression parameters to avoid power jumps. A joint loss function is used to optimize the output of each branch of the decoder, and the output parameters of each branch are aggregated to form an initial multi-source control parameter set containing the basic heating curve, thermal hysteresis compensation, and power switching jitter suppression. Through the above chain processing method, the composite fingerprint vector and recipe identifier are transformed into multi-source control parameters with execution significance, so as to achieve precise cooking mode matching under dynamic energy input conditions.

[0192] S8.5: Perform boundary constraint verification and timing smoothing filtering on the initial multi-source control parameter set to eliminate execution abrupt changes caused by model jumps, and finally generate the final control command containing thermal hysteresis compensation and power switching jitter suppression parameters and send it to the actuator.

[0193] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0194] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0195] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for mapping multiple energy modes of an oven to cooking parameters, specifically including: S1: Acquire timing characteristic signals such as voltage fluctuation spectrum distribution, power ramp-up slope envelope, start-stop transient response delay, and thermal inertia decay time constant under multiple energy input conditions; S2: Normalize the time-series feature signal to generate a high-dimensional vector as an energy fingerprint; S3: Input the energy fingerprint and the user-selected recipe identifier into the main track model to output a set of cooking parameters; S4: Collect the cavity temperature field curve, thermal image sequence and energy consumption trajectory after the control is executed, and compare the above real operation data with the cooking parameter set in time and space to generate deviation identification results; S5: Based on the deviation recognition result, determine whether the prediction error under the same type of energy fingerprint input for a consecutive preset number of times exceeds the set threshold. If it exceeds the threshold, trigger the fine-tuning command. S6: In response to the fine-tuning instruction, freeze the subset of neurons in the main track model that are strongly correlated with specific parameters, and use the current batch running data to perform single-step gradient updates to generate incremental patches; S7: Establish an index mapping relationship between fingerprints and patches, and store the incremental patches in the calibration cache so that they can be automatically loaded and superimposed on the main rail output when the same energy fingerprint appears in the future, forming an adaptively corrected combination of cooking parameters; S8: For multi-energy mixed input scenarios, the energy fingerprints of multiple independent energy sources are weighted and fused according to the real-time power ratio to generate a composite fingerprint, and the final control command is output by combining the main track model with thermal hysteresis compensation and power switching jitter suppression algorithm.

2. The method for mapping multiple energy modes of an oven to cooking parameters according to claim 1, characterized in that, Step S1 specifically includes: The original voltage and current time-series signals from multiple energy input terminals are acquired using an embedded high-speed data acquisition interface. The original voltage and current time-series signals are then decomposed in the frequency domain using a fast Fourier transform algorithm to extract voltage fluctuation spectrum distribution data. Based on the time-domain window corresponding to the voltage fluctuation spectrum distribution data, the slope of the original power sequence is calculated using a sliding differential filtering algorithm to extract the power ramp-up slope envelope data that reflects the thermal response speed of the heating element or the ignition efficiency of the burner. Based on the starting trigger point of the power ramp slope envelope data, the zero-crossing detection and threshold comparison logic is used to perform timing alignment analysis on the energy start-stop control command and the actual load response waveform, and the start-stop transient response delay data is extracted. Based on the cooling phase after the start-stop transient response delay data ends, the attenuation model matching process of the cavity temperature drop curve is performed using an exponential fitting regression algorithm to extract the thermal inertia decay time constant data. Based on the voltage fluctuation spectrum distribution data, power ramp slope envelope data, start-stop transient response delay data, and thermal inertia decay time constant data, the above-mentioned multidimensional heterogeneous physical quantities are aggregated into the time-series characteristic signal in a unified format using a multi-channel data fusion encapsulation protocol, which serves as the standard input object for subsequent normalization processing.

3. The method for mapping multiple energy modes of an oven to cooking parameters according to claim 1, characterized in that, Step S5 specifically includes: The single prediction error data in the deviation identification results are obtained, and the single prediction error data corresponding to the same type of energy fingerprint input for a consecutive preset number of times are processed by time sequence arrangement using the sliding time window algorithm to generate a continuous error data queue containing historical error sequences. Based on the historical error sequence in the continuous error data queue, a weighted moving average filtering algorithm is used to suppress noise and extract trends from the historical error sequence, generating a cumulative error feature vector. Obtain the error components of each dimension in the cumulative error feature vector, and use the dynamic threshold calculation strategy to adaptively correct the preset basic threshold based on the heat sensitivity coefficient of the current recipe identifier and the fluctuation intensity index of the energy fingerprint, so as to generate a real-time judgment threshold boundary for the current working condition. Based on the cumulative error feature vector and the real-time judgment threshold boundary, a multi-dimensional vector comparison operation is performed to compare the magnitude of the cumulative error feature vector with the real-time judgment threshold boundary to generate an error over-limit Boolean flag indicating whether the prediction accuracy has failed. In response to the true state of the error exceeding the limit Boolean flag, the local optimization process is activated when the error exceeding the limit Boolean flag is true using state machine transition logic, so as to generate the fine-tuning instruction that drives the main track model to update the weights.

4. The method according to claim 1, characterized in that, The high-dimensional vector form of the energy fingerprint is subjected to fixed-length binary encoding, padding or truncation, and cyclic redundancy check to generate a unique index number, which is then bound to the locally stored incremental patch file.

5. The method according to claim 1, characterized in that, When responding to the fine-tuning instruction, only the subset of neurons in the main track model that are strongly correlated with the prediction bias parameter are allowed to update their weights, while the remaining weights are frozen. Through single-step gradient updates and weight differencing, sparse and compressed incremental patches are generated.

6. The method according to claim 1, characterized in that, Based on the Least Recently Used (LRU) strategy, the index mapping relationship between energy fingerprints and incremental patches in the calibration cache is dynamically managed, and old incremental patches that are accessed infrequently are automatically released.

7. The method according to claim 1, characterized in that, In the multi-energy mixed input scenario, the real-time power values ​​of each independent energy source are weighted, and the energy fingerprints of each independent energy source are weighted and summed to form a composite fingerprint.

8. The method according to claim 7, characterized in that, The composite fingerprint is used as an index to retrieve data from the calibration cache. If an incremental patch is found, it is automatically loaded and superimposed on the output weights of the main track model to form a multi-source parameter output for dynamic calibration.

9. The method according to claim 7 or 8, characterized in that, The parameter set output by the main rail model includes the basic heating curve, thermal hysteresis compensation, and power switching jitter suppression, which are then sent to the oven control actuator through boundary constraints and timing filtering.

10. The method according to claim 1, characterized in that, The magnitude of the cumulative error vector is compared with the Euclidean norm of the dynamic threshold, and a third-order time continuous judgment logic is used to determine whether the prediction error under a preset number of consecutive times exceeds the set threshold.

Images