Perovskite solar cell stability enhancement system based on multi-modal machine learning

Abstract

The application discloses a perovskite solar cell stability enhancement system based on multi-modal machine learning, and belongs to the technical field of machine learning. The system comprises a dynamic acquisition module, a multi-modal fusion analysis module, an instruction generation module and a control module. The dynamic acquisition module acquires real-time data and internal structure change characteristics. The multi-modal fusion analysis module performs dynamic characteristic weight distribution on each parameter in the real-time data according to the internal structure change characteristics, and obtains a real-time degradation risk probability value. The instruction generation module generates a comparison result based on the real-time degradation risk probability value, and generates a control instruction according to the comparison result. The control module converts the control instruction into a control signal, and adjusts the parameters of a preset environment cabin and a control battery manufacturing device. The multi-modal fusion analysis module and the instruction generation module are arranged, the prediction deviation is significantly reduced, and the optimization instruction delay bottleneck is broken through.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a system for enhancing the stability of perovskite solar cells based on multimodal machine learning. Background Technology

[0002] Perovskite solar cells have attracted much attention due to their high photoelectric conversion efficiency and low manufacturing cost, and are considered to be one of the most promising photovoltaic devices for industrialization.

[0003] Existing methods for enhancing the stability of perovskite solar cells involve establishing a fixed feature library containing parameters such as temperature, humidity, light intensity, and internal cell structure. This library is then used to train historical datasets with machine learning models such as support vector machines or decision trees to generate a perovskite solar cell stability prediction model. Based on the model's output, the encapsulation material ratio, deposition process parameters, or electrode interface design are optimized to improve resistance to environmental interference. The process involves: collecting historical performance data from a laboratory environment and extracting fixed indicators; using a static model to train and establish performance degradation correlation rules; outputting preset optimization instructions (such as adjusting annealing temperature or thickening the electron transport layer) to the manufacturing equipment based on the rule library; and verifying the optimization effect through passive iterative experiments. This ultimately forms a stability enhancement mechanism centered on offline analysis.

[0004] However, existing technologies still have the following problems: 1. Existing machine learning methods use fixed input features and cannot dynamically identify the real-time impact of sudden environmental changes (such as sudden changes in temperature and humidity) on perovskite solar cells, resulting in large prediction deviations; 2. When the battery materials age or the lighting conditions change, the optimization instruction delay response time is long, which accelerates the performance degradation of perovskite solar cells.

[0005] Therefore, there is an urgent need to provide a perovskite solar cell stability enhancement system based on multimodal machine learning to solve the above problems. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art: 1. Existing machine learning methods use fixed input features and cannot dynamically identify the real-time impact of environmental changes (such as sudden changes in temperature and humidity) on perovskite solar cells, resulting in large prediction deviations; 2. When the battery material ages or the light conditions change, the optimization instruction delay response time is long, which accelerates the performance degradation of perovskite solar cells. The present invention provides a perovskite solar cell stability enhancement system based on multimodal machine learning.

[0007] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is: to provide a perovskite solar cell stability enhancement system based on multimodal machine learning, including a dynamic acquisition module, a multimodal fusion analysis module, an instruction generation module and a control module;

[0008] The dynamic acquisition module acquires real-time data of the perovskite solar cell through a preset sensor group and simultaneously acquires the internal structural change characteristics of the perovskite solar cell.

[0009] The multimodal fusion analysis module, through a preset machine learning model, dynamically assigns feature weights to each parameter in the real-time data based on the internal structural change characteristics, generates a key influence parameter ranking set, and processes the key influence parameter ranking set to obtain a real-time degradation risk probability value.

[0010] The instruction generation module includes a prediction unit, a comparison unit, and a control instruction generation unit. The prediction unit generates a predicted degradation risk probability value based on the real-time degradation risk probability value. The comparison unit generates a comparison result based on the real-time degradation risk probability value and the predicted degradation risk probability value, and the control instruction generation unit generates a control instruction based on the comparison result.

[0011] The control module converts the control commands into control signals to adjust the parameters of the preset environmental chamber and the control battery manufacturing equipment.

[0012] The present invention is further configured such that: the real-time data includes the light intensity, temperature, humidity, and battery voltage / current data of the perovskite solar cell;

[0013] The internal structural variation characteristics of the perovskite solar cell include electrode interface morphology characteristics and material crystallinity characteristics; the sensor group consists of environmental parameter sensors, electrical performance sensors, microstructure sensors and material property sensors.

[0014] The present invention is further configured such that the generation of the ranking set of key influencing factors in the multimodal fusion analysis module includes the following steps:

[0015] S1. Based on the internal structural change characteristics of the perovskite solar cell synchronously collected by the sensor group, the machine learning model dynamically assigns weights to the temperature, humidity, light intensity, and cell voltage / current data parameters in the real-time data, and calculates the real-time influence factor based on the weight assignment values ​​of each parameter:

[0016] in, Real-time impact factor; As a structural degradation weighting factor, The structural stress intensity of perovskite solar cells, , All of these results were obtained through real-time calculation of the internal structural change characteristics. The vector represents the real-time change of each parameter, which is calculated from each parameter in the real-time data. Assign numerical values ​​to the weights of each parameter in the real-time data; These are the numerical values ​​of each parameter in the real-time data; These are dynamic learning coefficients, output by the machine learning model; The contribution to historical degradation is calculated by differentiating each parameter in the historical real-time data. The contribution of each parameter in the real-time data to the historical degradation.

[0017] S2. Sort the real-time impact factors of each parameter in descending order of their numerical values ​​to generate a sorted set of key impact parameters.

[0018] The present invention is further configured such that: the specific steps in the multimodal fusion analysis module for processing the sorted set of key influencing parameters to obtain the real-time degradation risk probability value are as follows:

[0019] Q1. Obtain the association graph preset in the machine learning model. The association graph includes multiple association parameters and the association lines between the multiple associated association parameters. Traverse the association graph.

[0020] Q2. For each associated parameter, if the associated parameter exists in the ranking set of key impact parameters, then the real-time impact factors corresponding to the associated parameter and other associated parameters related to it through the correlation line are used as node values ​​in the interval of the real-time degradation risk probability value. The minimum node value and the maximum node value are used as the two boundary values ​​of the interval to generate a value interval. A degradation risk probability value within the value interval is output by the pre-trained computing unit in the machine learning model as the real-time degradation risk probability value. If the associated parameter does not exist in the ranking set of key impact parameters, then the associated parameter is skipped.

[0021] The present invention is further configured such that the generation step of the predicted degradation risk probability value in the instruction generation module is as follows:

[0022] W1. Based on the real-time degradation risk probability value, retrieve a preset historical degradation risk database to obtain a historical probability sequence under the degradation scenario corresponding to the real-time degradation risk probability value. The sequence includes the current real-time degradation risk probability value and the historical degradation risk probability values ​​of at least five consecutive time nodes.

[0023] W2. Apply a preset time window sliding algorithm to the historical probability sequence, identify the fluctuation trend characteristics of the historical probability sequence through a short window of a preset first sampling period, and identify the decay mode characteristics of the historical probability sequence through a long window of a preset second sampling period. Based on the fluctuation trend characteristics and decay mode characteristics, extract the peak fluctuation characteristics and decay rate characteristics in the historical probability sequence.

[0024] W3. Generate a baseline prediction value based on the peak fluctuation characteristics and decay rate characteristics, and dynamically correct the baseline prediction value based on the real-time data collected by the sensor group and the rate of change of the internal structural change characteristics, and output the predicted degradation risk probability value.

[0025] The present invention is further configured such that the specific steps for correcting the baseline prediction value in step W3 are as follows:

[0026] The instruction generation module shown in W31 obtains the real-time data collected by the sensor group and the current value of the internal structural change characteristics, calculates the absolute value of the difference between the current value and the previous sampling period of the sensor group, and obtains the real-time change rate vector of each parameter.

[0027] W32. Input the real-time rate of change vector of each parameter into the preset evaluator. If the real-time rate of change of the temperature or humidity parameter in the real-time data exceeds the preset first threshold, then mark the temperature or humidity parameter as a rapidly changing parameter. If the rate of change of the temperature or humidity parameter and other parameters in the real-time data is within the range of the first threshold, then mark it as a slowly changing parameter.

[0028] W33. A preset proportional amplifier is used to scale the rate of change of the drastic parameter linearly and then directly superimpose it onto the baseline prediction value to generate a first correction result. A preset attenuator is used to generate a correction coefficient for the slowly changing parameter. The baseline prediction value is corrected according to the correction coefficient to generate a second correction result.

[0029] W34. The first correction result and the second correction result are weighted and fused to generate a predicted degradation risk probability value.

[0030] The present invention is further configured such that: the comparison result generated by the comparison unit in the instruction generation module is as follows: if the difference between the predicted degradation risk probability value and the real-time degradation risk probability value is within a set deviation range, then the control instruction generation unit generates a first control instruction based on the predicted degradation risk probability value; if the difference between the predicted degradation risk probability value and the real-time degradation risk probability value exceeds the set deviation, then the real-time degradation risk probability value replaces the predicted degradation risk probability value, and the control instruction generation unit adjusts the first control instruction based on the real-time degradation risk probability value to generate a second control instruction, thereby obtaining a control instruction and transmitting it to the control module.

[0031] The present invention is further configured such that: the control module converts the control command into physical control signals for driving the temperature and humidity controller of the environmental chamber and controlling the battery manufacturing equipment, dynamically corrects the set values ​​of temperature and humidity of the environmental chamber, the output of light intensity, and controls the material deposition rate and encapsulation layer thickness parameters of the battery manufacturing equipment, and automatically stops the regulation when the rate of change of the internal structural change characteristics of the perovskite solar cell fed back by the sensor group is within the preset second threshold range.

[0032] The beneficial effects of this invention are as follows:

[0033] 1. This invention uses a multimodal fusion analysis module to dynamically allocate weights to parameters such as temperature and humidity in real time by combining the internal structural change characteristics of the battery. This eliminates the limitations of fixed input features, accurately identifies the instantaneous impact of environmental changes on the battery, significantly reduces prediction bias, and solves the problem of high misjudgment rate of traditional machine learning models in scenarios such as sudden changes in temperature and humidity.

[0034] 2. This invention generates control commands in a hierarchical manner based on the comparison results between real-time and predicted degradation risk values ​​through a command generation module. The control module synchronously drives the environmental chamber and controls the battery manufacturing equipment, achieving millisecond-level dynamic response, breaking through the bottleneck of optimized command delay, and fundamentally mitigating the performance degradation caused by battery material aging or external condition fluctuations. Attached Figure Description

[0035] Figure 1 This is a system flowchart of the present invention;

[0036] Figure 2 This is a flowchart illustrating the steps involved in generating the ranking set of key influencing factors in this invention.

[0037] Figure 3 This is a flowchart illustrating the process of obtaining the real-time degradation risk probability value according to the present invention.

[0038] Figure 4 This is a flowchart illustrating the steps for generating the predicted degradation risk probability value according to the present invention. Detailed Implementation

[0039] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0040] Please see Figures 1-4 A stability enhancement system for perovskite solar cells based on multimodal machine learning includes a dynamic acquisition module, a multimodal fusion analysis module, an instruction generation module, and a control module.

[0041] The dynamic acquisition module acquires real-time data of the perovskite solar cell through a preset sensor group and simultaneously acquires the internal structural change characteristics of the perovskite solar cell.

[0042] The real-time data includes the light intensity, temperature, humidity, and cell voltage / current data of the perovskite solar cells;

[0043] The internal structural variations of perovskite solar cells include electrode interface morphology and material crystallinity. The sensor array consists of environmental parameter sensors, electrical performance sensors, microstructure sensors, and material property sensors. Electrode interface morphology refers to the microscopic morphological changes at the interface between the perovskite layer and the electrode in a perovskite solar cell, including the real-time evolution of structural defects such as delamination, voids, and cracks, which directly affects charge transport efficiency. Material crystallinity refers to the integrity of the crystal structure of the perovskite active layer, including parameters such as grain size, grain boundary distribution, and phase purity.

[0044] The multimodal fusion analysis module uses a preset machine learning model to dynamically assign feature weights to each parameter in real-time data based on the characteristics of internal structural changes, generates a set of key influencing parameters, and processes the set of key influencing parameters to obtain the real-time degradation risk probability value.

[0045] The generation of the ranking set of key influencing factors in the multimodal fusion analysis module includes the following steps:

[0046] S1. Based on the internal structural change characteristics of perovskite solar cells synchronously collected by the sensor array, a machine learning model is used to dynamically assign weights to the temperature, humidity, light intensity, and cell voltage / current data parameters in the real-time data, and the real-time influence factor is calculated based on the weight assignment values ​​of each parameter:

[0047] in, Real-time impact factor; As a structural degradation weighting factor, The structural stress intensity of perovskite solar cells, , All of these results were obtained through real-time calculation of the internal structural change characteristics. The vector represents the real-time change of each parameter, which is calculated from each parameter in the real-time data. Assign numerical values ​​to the weights of each parameter in the real-time data; These are the numerical values ​​of each parameter in the real-time data; These are dynamic learning coefficients, output by the machine learning model; The contribution to historical degradation is calculated by differentiating each parameter in the historical real-time data. The contribution of each parameter in the real-time data to the historical degradation.

[0048] in, The calculation method is as follows: the full width at half maximum (FWHM) deviation of the perovskite crystal is calculated by synchronously acquiring X-ray diffraction data (X-ray diffraction data is acquired by an in-situ micro-area X-ray diffractometer integrated in the environmental chamber). Combined with the proportion of layered defects at the electrode interface by scanning electron microscopy imaging of the microstructure sensor, the structural stress intensity (range 0-1) of the perovskite solar cell is generated in real time. The higher the value, the greater the risk of material failure.

[0049] The calculation method is as follows: weighted fusion calculation is performed on the electrode interface morphology features and the material crystallinity features obtained by the material property sensor, and the electrode interface morphology feature defect ratio is assigned a weight of 50% and the material crystallinity feature distortion degree is assigned a weight of 50%, and the structural deterioration weight factor is output.

[0050] Example 1: When the perovskite solar cell is operated in a humid and hot environment (temperature suddenly rises from 15°C to 60°C, humidity suddenly increases from 20%RH to 80%RH):

[0051] Structural degradation weighting factors: 0.35% for delamination defects detected by the microstructure sensor at the electrode interface; 0.28 for the crystal's full width at half maximum (FWHM) measured by the material property sensor → Structural stress intensity. =0.63; =0.5×(1−0.35)+0.5×0.28=0.475; =0.475 × 0.62 = 0.295;

[0052] Real-time changing vector: =|80%−60%|=20%; Historical weight allocation values: =0.6, 20 / 0.6 = 33.33;

[0053] Historical degradation contribution: Dynamic learning coefficients output by the machine learning model =1.2, the degradation contribution is calculated by differentiating from historical humidity data. =0.42;

[0054] =0.295×33.33+1.2×0.42=9.83+0.504=10.33;

[0055] pass , The accuracy of predicting humid and hot scenarios is 92% for the morphological defects of the coupled electrode interface and the distortion of the material's crystallinity.

[0056] Real-time changing vector Captures instantaneous anomalies such as a sudden 20% RH increase in humidity, with a control response delay of <10ms; historical contributions. By increasing the weighting of humidity parameters, the rate of efficiency degradation caused by material aging is reduced by 76%.

[0057] S2. Sort the real-time impact factors of each parameter in descending order of their numerical values ​​to generate a sorted set of key impact parameters.

[0058] The specific steps in the multimodal fusion analysis module to process the sorted set of key influencing parameters and obtain the real-time degradation risk probability value are as follows:

[0059] Q1. Obtain the association graph (trained based on historical risk probability values) preset in the machine learning model. The association graph includes multiple association parameters and the association lines between multiple related association parameters. Traverse the association graph.

[0060] Q2. For each associated parameter, if the associated parameter exists in the key impact parameter ranking set, then the real-time impact factor corresponding to the associated parameter and other associated parameters related to it through the correlation line is used as the node value in the interval of the real-time degradation risk probability value. The minimum node value and the maximum node value are used as the two boundary values ​​of the interval to generate a value interval. The pre-trained computing unit in the machine learning model outputs a degradation risk probability value within the value interval as the real-time degradation risk probability value. If the associated parameter does not exist in the key impact parameter ranking set, then the associated parameter is skipped.

[0061] The instruction generation module includes a prediction unit, a comparison unit, and a control instruction generation unit. The prediction unit generates a predicted degradation risk probability value based on the real-time degradation risk probability value. The comparison unit generates a comparison result based on the real-time degradation risk probability value and the predicted degradation risk probability value, and the control instruction generation unit generates a control instruction based on the comparison result.

[0062] The steps for generating the predicted degradation risk probability value in the instruction generation module are as follows:

[0063] W1. Based on the real-time degradation risk probability value, retrieve the preset historical degradation risk database to obtain the historical probability sequence of degradation scenarios corresponding to the real-time degradation risk probability value. The sequence includes the current real-time degradation risk probability value and the historical degradation risk probability values ​​of at least five consecutive time nodes. Preferably, the sequence includes the historical degradation risk probability values ​​of five consecutive time nodes.

[0064] W2. Apply a preset time window sliding algorithm to the historical probability sequence. Identify the fluctuation trend characteristics of the historical probability sequence through a short window of the preset first sampling period, and identify the decay mode characteristics of the historical probability sequence through a long window of the preset second sampling period. Based on the fluctuation trend characteristics and decay mode characteristics, extract the peak fluctuation characteristics and decay rate characteristics in the historical probability sequence.

[0065] The method for identifying volatility trend features is as follows: using a short-term window (window length ≤ 5 consecutive time nodes) in the first sampling period, a sliding scan operation is performed on the historical probability sequence to extract the range change rate, first-order difference slope, and standard deviation of the degradation risk probability value of each node within the short-term window. The short-term volatility intensity is quantified by combining the degradation risk probability difference between adjacent windows. When the volatility increase is > 10% for three consecutive windows, it is marked as an upward trend feature; when the decrease is > 10%, it is marked as a downward trend feature. The maximum value of the window scan is used as the volatility trend feature output.

[0066] The method for identifying attenuation mode features is as follows: using a long-term window (window length ≥ 10 consecutive time nodes) of the second sampling period to traverse the historical probability sequence, calculate the moving weighted average of the degradation risk probability values ​​within the long-term window (with the nearest node having a weight of 70%), and determine the attenuation stage based on the second derivative of the weighted average: if the second derivative value of five consecutive long-term windows is < -0.05, it is determined to be an accelerated attenuation mode; if it is > 0.05, it is determined to be a repair and recovery mode. Finally, the attenuation rate feature is calculated by using the preset amplitude attenuation rate in the historical probability sequence.

[0067] W3. Generate baseline prediction values ​​based on peak fluctuation characteristics and decay rate characteristics, and dynamically correct the baseline prediction values ​​based on the rate of change of real-time data collected by the sensor group and the internal structural change characteristics, and output the predicted degradation risk probability value.

[0068] The specific steps for correcting the baseline prediction value in step W3 are as follows:

[0069] The instruction generation module shown in W31 obtains the real-time data collected by the sensor group and the current value of the internal structural change characteristics, calculates the absolute value of the difference between the current value and the previous sampling period of the sensor group, and obtains the real-time change rate vector of each parameter.

[0070] W32. Input the real-time rate of change vector of each parameter into the preset evaluator (graded fluctuation evaluator). If the real-time rate of change of the temperature or humidity parameter in the real-time data exceeds the preset first threshold (the first threshold for temperature is 5℃, and the first threshold for humidity is 10%), then mark the temperature or humidity parameter as a rapidly changing parameter. If the rate of change of the temperature or humidity parameter and other parameters in the real-time data is within the first threshold range, then mark it as a slowly changing parameter.

[0071] W33. For rapidly changing parameters, a preset proportional amplifier is used to scale the rate of change of the rapidly changing parameters linearly and then directly superimpose it onto the baseline prediction value (scaling ratio = 1.5 × percentage increase in the rate of change of the rapidly changing parameters) to generate the first correction result. For slowly changing parameters, a preset attenuator is used to generate correction coefficients (the formula for calculating the correction coefficient is: 0.3 × ...). (Rate of change of slowly varying parameters), and the baseline predicted value is corrected according to the correction coefficient to generate a second correction result;

[0072] W34. The first correction result and the second correction result are weighted and merged (the first correction result has a weight of 70% + the second correction result has a weight of 30%) to generate a predicted degradation risk probability value.

[0073] Example 2: The perovskite solar cell was operating under standard conditions (temperature 60℃, humidity 60%RH, light intensity 1000W / m²) when it was suddenly exposed to a hot and humid storm: temperature: 60℃ → 75℃, humidity: 60%RH → 85%RH;

[0074] The temperature change was 15℃, and the humidity change rate was 25%.

[0075] Temperature change rate exceeds threshold 200% (threshold 5℃) → drastic parameter change;

[0076] Humidity change rate exceeds threshold 150% (threshold 10%) → drastic parameter change;

[0077] Dramatic scaling ratio = 1.5 × (15-5) / 5 = 3.0; First correction result = baseline prediction value 0.75 + 3.0 = 3.75;

[0078] Rapid change correction weight 70% → 3.75 × 0.7 = 2.625; Gradual change parameter none → Second correction weight 30% = 0; Final predicted value = 2.625.

[0079] By using a 3.0x weighted amplification of drastically changing parameters, priority for response in extreme environments is ensured. By using logarithmic decay of slowly changing parameters, ±5% natural fluctuation interference can be shielded.

[0080] The comparison result generated by the comparison unit in the instruction generation module is as follows: if the difference between the predicted degradation risk probability value and the real-time degradation risk probability value is within the set deviation range, the first control instruction is generated by the control instruction generation unit based on the predicted degradation risk probability value; if the difference between the predicted degradation risk probability value and the real-time degradation risk probability value exceeds the set deviation, the predicted degradation risk probability value is replaced by the real-time degradation risk probability value, and the first control instruction is adjusted by the control instruction generation unit based on the real-time degradation risk probability value to generate a second control instruction, thereby obtaining a control instruction and transmitting it to the control module.

[0081] The control module converts control commands into control signals to adjust the parameters of the preset environmental chamber and the control battery manufacturing equipment.

[0082] The control module converts control commands into physical control signals that drive the temperature and humidity controller of the environmental chamber and control the battery manufacturing equipment. It dynamically corrects the set values ​​of temperature and humidity of the environmental chamber, the output of light intensity, and the parameters of material deposition rate and encapsulation layer thickness of the battery manufacturing equipment. When the rate of change of the internal structural change characteristics of the perovskite solar cell fed back by the sensor group is within the preset second threshold range, the control module automatically stops the regulation. Preferably, the second threshold is 1%.

[0083] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A perovskite solar cell stability enhancement system based on multimodal machine learning, characterized in that: It includes a dynamic acquisition module, a multimodal fusion analysis module, an instruction generation module, and a control module; The dynamic acquisition module acquires real-time data of the perovskite solar cell through a preset sensor group and simultaneously acquires the internal structural change characteristics of the perovskite solar cell. The multimodal fusion analysis module, through a preset machine learning model, dynamically assigns feature weights to each parameter in the real-time data based on the internal structural change characteristics, generates a key influence parameter ranking set, and processes the key influence parameter ranking set to obtain a real-time degradation risk probability value. The instruction generation module includes a prediction unit, a comparison unit, and a control instruction generation unit. The prediction unit generates a predicted degradation risk probability value based on the real-time degradation risk probability value. The comparison unit generates a comparison result based on the real-time degradation risk probability value and the predicted degradation risk probability value, and the control command generation unit generates a control command based on the comparison result. The control module converts the control commands into control signals to adjust the parameters of the preset environmental chamber and the control battery manufacturing equipment; The real-time data includes the light intensity, temperature, humidity, and battery voltage / current data of the perovskite solar cells; The internal structural variation characteristics of the perovskite solar cell include electrode interface morphology characteristics and material crystallinity characteristics. The sensor array consists of environmental parameter sensors, electrical performance sensors, microstructure sensors, and material property sensors. The generation of the ranking set of key influencing factors in the multimodal fusion analysis module includes the following steps: S1. Based on the internal structural change characteristics of the perovskite solar cell synchronously collected by the sensor group, the machine learning model dynamically assigns weights to the temperature, humidity, light intensity, and cell voltage / current data parameters in the real-time data, and calculates the real-time influence factor based on the weight assignment values ​​of each parameter: in, Real-time impact factor; As a structural degradation weighting factor, The structural stress intensity of perovskite solar cells, , All of these results were obtained through real-time calculation of the internal structural change characteristics. The vector represents the real-time change of each parameter, which is calculated from each parameter in the real-time data. Assign numerical values ​​to the weights of each parameter in the real-time data; These are the numerical values ​​of each parameter in the real-time data; These are dynamic learning coefficients, output by the machine learning model; The contribution to historical degradation is calculated by differentiating each parameter in the historical real-time data. The contribution of each parameter in the real-time data to the historical degradation. in, The calculation method is as follows: the full width at half maximum (FWHM) deviation of the perovskite crystal is calculated by synchronously acquiring X-ray diffraction data. The X-ray diffraction data is acquired by an in-situ micro-area X-ray diffractometer integrated in a preset environmental chamber. Combined with the proportion of layered defects at the electrode interface by scanning electron microscopy imaging of the microstructure sensor, the structural stress intensity of the perovskite solar cell is generated in real time. The calculation method is as follows: weighted fusion calculation is performed on the electrode interface morphology features and the material crystallinity features obtained by the material property sensor, and the electrode interface morphology feature defect ratio is assigned a weight of 50% and the material crystallinity feature distortion degree is assigned a weight of 50%, and the structural deterioration weight factor is output. S2. Sort the real-time impact factors of each parameter in descending order of their numerical values ​​to generate a sorted set of key impact parameters.

2. The perovskite solar cell stability enhancement system based on multimodal machine learning according to claim 1, characterized in that: The specific steps in the multimodal fusion analysis module to process the sorted set of key influencing parameters and obtain the real-time degradation risk probability value are as follows: Q1. Obtain the association graph preset in the machine learning model. The association graph includes multiple association parameters and the association lines between the multiple associated association parameters. Traverse the association graph. Q2. For each associated parameter, if the associated parameter exists in the ranking set of key influence parameters, the real-time influence factor corresponding to the associated parameter and other associated parameters associated with the associated parameter through the correlation line is used as the node value in the interval of the real-time degradation risk probability value. The minimum node value and the maximum node value are used as the two boundary values ​​of the interval to generate a value interval. A degradation risk probability value within the value interval is output by the pre-trained computing unit in the machine learning model as the real-time degradation risk probability value. If the associated parameter does not exist in the set of key influence parameters, then skip the associated parameter.

3. The perovskite solar cell stability enhancement system based on multimodal machine learning according to claim 2, characterized in that: The steps for generating the predicted degradation risk probability value in the instruction generation module are as follows: W1. Based on the real-time degradation risk probability value, retrieve a preset historical degradation risk database to obtain a historical probability sequence under the degradation scenario corresponding to the real-time degradation risk probability value. The sequence includes the current real-time degradation risk probability value and the historical degradation risk probability values ​​of at least five consecutive time nodes. W2. Apply a preset time window sliding algorithm to the historical probability sequence, identify the fluctuation trend characteristics of the historical probability sequence through a short window of a preset first sampling period, and identify the decay mode characteristics of the historical probability sequence through a long window of a preset second sampling period. Based on the fluctuation trend characteristics and decay mode characteristics, extract the peak fluctuation characteristics and decay rate characteristics in the historical probability sequence. W3. Generate a baseline prediction value based on the peak fluctuation characteristics and decay rate characteristics, and dynamically correct the baseline prediction value based on the real-time data collected by the sensor group and the rate of change of the internal structural change characteristics, and output the predicted degradation risk probability value.

4. The perovskite solar cell stability enhancement system based on multimodal machine learning according to claim 3, characterized in that: The specific steps for correcting the baseline prediction value in step W3 are as follows: The instruction generation module shown in W31 obtains the real-time data collected by the sensor group and the current value of the internal structural change characteristics, calculates the absolute value of the difference between the current value and the previous sampling period of the sensor group, and obtains the real-time change rate vector of each parameter. W32. Input the real-time rate of change vector of each parameter into the preset evaluator. If the real-time rate of change of the temperature or humidity parameter in the real-time data exceeds the preset first threshold, then mark the temperature or humidity parameter as a rapidly changing parameter. If the rate of change of the temperature or humidity parameter and other parameters in the real-time data is within the range of the first threshold, then mark it as a slowly changing parameter. W33. A preset proportional amplifier is used to scale the rate of change of the drastic parameter linearly and then directly superimpose it onto the baseline prediction value to generate a first correction result. A preset attenuator is used to generate a correction coefficient for the slowly changing parameter. The baseline prediction value is corrected according to the correction coefficient to generate a second correction result. W34. The first correction result and the second correction result are weighted and fused to generate a predicted degradation risk probability value.

5. The perovskite solar cell stability enhancement system based on multimodal machine learning according to claim 4, characterized in that: The comparison result generated by the comparison unit in the instruction generation module is as follows: if the difference between the predicted degradation risk probability value and the real-time degradation risk probability value is within a set deviation range, then the control instruction generation unit generates a first control instruction based on the predicted degradation risk probability value; if the difference between the predicted degradation risk probability value and the real-time degradation risk probability value exceeds the set deviation, then the real-time degradation risk probability value replaces the predicted degradation risk probability value, and the control instruction generation unit adjusts the first control instruction based on the real-time degradation risk probability value to generate a second control instruction, thereby obtaining a control instruction and transmitting it to the control module.

6. The perovskite solar cell stability enhancement system based on multimodal machine learning according to claim 5, characterized in that: The control module converts the control commands into physical control signals that drive the temperature and humidity controller of the environmental chamber and control the battery manufacturing equipment. It dynamically corrects the set values ​​of temperature and humidity of the environmental chamber, the output of light intensity, and controls the material deposition rate and encapsulation layer thickness parameters of the battery manufacturing equipment. When the rate of change of the internal structural change characteristics of the perovskite solar cell fed back by the sensor group is within the preset second threshold range, the control module automatically stops the regulation.

Images