Battery production waste combustion risk intelligent early warning and disposal system

Abstract

The invention relates to the technical field of battery waste management and control, in particular to a battery production waste combustion risk intelligent early warning and disposal system. The system comprises a data acquisition module, a risk prediction module, an early warning management module, a disposal management module, a feedback optimization module and a parameter optimization module. The data acquisition module acquires waste multi-source monitoring data within a first preset time period; the risk prediction module trains a combustion risk prediction model accordingly; the early warning management module obtains real-time monitoring data in a second preset time period and executes early warning operation; the disposal management module executes disposal operation in a third preset time period according to the early warning result data; the feedback optimization module optimizes the prediction model in combination with early warning and disposal results; and the parameter optimization module configures an optimization target based on historical data and executes parameter directional optimization. The system can realize comprehensive monitoring, accurate early warning, effective disposal and dynamic optimization of the waste combustion risk.

Description

Technical Field

[0001] The present invention relates to the technical field of battery waste management and control, and in particular to an intelligent early warning and disposal system for combustion risks of battery production waste. Background Art

[0002] The battery production process generates a large amount of waste, including electrode materials, electrolyte residues, and separator fragments. Some of this waste is flammable and can pose a combustion risk during stacking, transportation, and temporary storage due to changes in environmental conditions or reactions within its components. With the rapid development of the new energy industry, battery production continues to rise, and the corresponding amount of production waste is also increasing. As waste storage scales expand, the probability of combustion risks increases. Currently, the management and control of the combustion risk of battery production waste mostly relies on manual inspections and traditional monitoring methods. Manual inspections have the problem of fixed inspection intervals and the inability to cover all waste storage areas in real time. Often, the combustion risk is not discovered until it has already appeared or even in the early stages of occurrence, missing the best time for disposal. Traditional monitoring methods are mostly limited to a single monitoring indicator, such as only monitoring the ambient temperature or smoke concentration, which makes it difficult to fully reflect the combustion risk status of battery production waste. Because the combustion process of battery production waste involves multiple links such as component decomposition, heat accumulation, and gas release, single indicator monitoring is prone to misjudgment or omission. Either normal environmental fluctuations are misjudged as combustion risks, resulting in unnecessary disposal operations and increased management and control costs; or subtle changes in the early stages of combustion risks are not captured in time, resulting in increased risks. In the existing risk management and control system, the early warning and disposal links are independent of each other, and there is a lack of effective data linkage and feedback mechanisms. When risk signals are detected, the transmission efficiency of early warning information is low, and the formulation of disposal plans often relies on the experience of staff. There is a lack of scientific guidance based on historical data and real-time risk status, and the targetedness and effectiveness of disposal measures are difficult to guarantee. In addition, over time, the composition of battery production waste may change, and new influencing factors may also appear in the external storage environment. The existing management and control methods are unable to dynamically optimize risk judgment standards and disposal strategies based on these changes, resulting in a gradual decline in management effectiveness and difficulty in adapting to the long-term and dynamic risk management needs of battery production waste combustion. Summary of the Invention

[0003] The main purpose of the present invention is to provide an intelligent early warning and disposal system for battery production waste combustion risks, aiming to solve the technical problems in the prior art.

[0004] The present invention proposes an intelligent early warning and disposal system for battery production waste combustion risks, comprising: a data acquisition module for acquiring multi-source monitoring data of battery production waste within a first preset time period; Risk prediction module, used to train combustion risk prediction models based on multi-source monitoring data; An early warning management module is used to obtain real-time monitoring data within a second preset time period and perform early warning operations on the real-time monitoring data; a disposal management module, configured to obtain warning result data within a third preset time period and perform disposal operations on the warning result data; Feedback optimization module, used to optimize the combustion risk prediction model based on the results of early warning operations and disposal operations; The parameter optimization module is used to configure risk optimization objectives based on historical warning data and historical disposal data and perform parameter-oriented optimization.

[0005] Preferably, the data acquisition module maps the multi-source monitoring data to a feature vector space after preprocessing to obtain a feature vector; Preprocessing includes data cleaning to remove noise data and outliers from multi-source monitoring data to obtain cleaned monitoring data; Mapping the cleaned monitoring data to the feature vector space to obtain the initial feature vector; The initial feature vector is convolved based on the convolutional neural network to obtain the feature vector.

[0006] Preferably, the risk prediction module is trained to obtain a combustion risk prediction model based on multi-source monitoring data and historical combustion event data; The combustion risk prediction model includes a risk generator and a risk discriminator; The feature vector is input into the risk generator to obtain the risk prediction value, and the historical combustion event data is input into the risk discriminator to make a true probability judgment based on the risk prediction value; If the absolute difference between the mean of the true probability and the preset convergence threshold for multiple consecutive times is less than or equal to the preset difference threshold, the training converges and the combustion risk prediction model is obtained.

[0007] Preferably, the early warning management module extracts a risk generator and a risk discriminator of a combustion risk prediction model; Based on the risk generator, early warning is performed on the real-time monitoring data to obtain the combustion risk warning value; Based on real-time monitoring data and combustion risk warning values, a rule mapping processing unit is constructed; The mapping relationship of the processing unit is mapped based on the reverse optimization rule of the risk discriminator.

[0008] Preferably, based on the real-time monitoring data and the combustion risk warning value, a rule mapping processing unit is constructed, including: After preprocessing the real-time monitoring data, feature extraction is performed on the preprocessed data using a feature extraction tool to obtain multiple feature fields; Extract multiple warning fields from the combustion risk warning value; Establish a mapping relationship between feature fields and warning fields, including: Initialize and generate a mapping scheme that meets the constraint conditions; wherein the mapping scheme includes a mapping relationship between each warning field and a number of feature fields; Constraints include that each warning field must have a mapping relationship with at least one feature field, each warning field must have a mapping relationship with at most all feature fields, and each warning field must have and can only have one mapping relationship with each feature field. A rule mapping processing unit is constructed based on the mapping scheme.

[0009] Preferably, the mapping relationship of the rule mapping processing unit is reversely optimized based on the risk discriminator, including determining the score of the rule mapping processing unit; Based on the rule mapping processing unit, a corresponding rule combustion risk warning value is generated for the real-time monitoring data; The risk discriminator based on the combustion risk prediction model determines the true probability of the rule combustion risk warning value; The absolute difference between the true probability and the preset convergence threshold is used as the score of the rule mapping processing unit; If the absolute difference between the score and the preset convergence threshold is greater than the preset difference threshold, the mapping relationship of the mapping scheme is updated based on the gradient descent method until the absolute difference between the score and the preset convergence threshold is less than or equal to the preset difference threshold.

[0010] Preferably, the handling management module classifies the warning result data into regular risk events and non-regular risk events; If it is a rule risk event, the warning result data is processed based on the rule mapping processing unit; If it is a non-regular risk event, the warning result data will be processed based on the risk generator.

[0011] Preferably, the warning result data is classified into regular risk events and non-regular risk events, including: If the absolute difference between the score of the rule mapping processing unit and the preset convergence threshold is less than or equal to the preset difference threshold, the corresponding real-time monitoring data is regarded as a rule risk event; Both real-time monitoring data and early warning result data are pre-processed and feature extracted; Based on the order of the characteristic fields in the real-time monitoring data and the early warning result data, the corresponding characteristic sequences are constructed respectively; Calculate the similarity value of the corresponding feature sequence; If the similarity value is greater than or equal to the preset similarity threshold, the warning result data is classified as a regular risk event; otherwise, the warning result data is classified as a non-regular risk event.

[0012] Preferably, calculating the similarity value of the corresponding feature sequence includes: Load the sliding extraction window to convert the feature sequence into a convolution sequence with a fixed dimension; Calculate the similarity value of the feature combination of each unit in the corresponding convolution sequence respectively; wherein the feature combination is extracted from the feature sequence based on the sliding extraction window; The similarity value is determined based on the average of the similarity values ​​of the feature combinations of all units.

[0013] Preferably, the feedback optimization module inputs the warning operation results and the disposal operation results as feedback information into the combustion risk prediction model to dynamically optimize the combustion risk prediction model; The parameter optimization module configures the risk optimization target based on the feedback information and performs parameter-oriented optimization. The parameter-oriented optimization result is used to update the training parameters of the risk prediction module.

[0014] The beneficial effects of the present invention are: By setting up a data acquisition module, multi-source monitoring data of battery production waste can be obtained within a first preset time period. Compared with the traditional single-indicator monitoring method, multi-source data can reflect the status of the waste from multiple dimensions, including but not limited to the temperature, humidity, and gas composition of the waste itself, as well as the temperature and oxygen concentration of the storage environment, etc., to achieve comprehensive capture of information related to combustion risks, avoid risk misjudgment or omission due to the one-sidedness of single data, and make risk monitoring more comprehensive and accurate. The risk prediction module trains a combustion risk prediction model based on multi-source monitoring data. This allows the model to fully learn the data characteristics of the battery production waste combustion risk formation process, more accurately identify the risk level in different situations, and provide reliable model support for subsequent early warning operations. Unlike the traditional risk assessment method that relies on human experience, the data-based training model can reduce the influence of human subjective factors and make risk assessment more objective. As data accumulates, the model's risk prediction ability will continue to improve, better adapting to the risk assessment needs of different batches and compositions of battery production waste. The early warning management module acquires real-time monitoring data and executes early warning operations within a second preset time period, enabling real-time monitoring and timely early warning of the risk of battery waste combustion. Real-time monitoring ensures that risk signals are quickly captured. Once real-time data indicates a risk, early warning operations can be quickly initiated, promptly transmitting risk information to relevant management personnel. This overcomes the time interval limitations of traditional manual inspections, avoids the escalation of risks due to information lags, and purchases more time for subsequent disposal operations.

[0015] The disposal management module acquires the warning result data and executes the disposal operation within the third preset time period, achieving an effective connection between warning and disposal. Warning result data can provide clear risk guidance for disposal operations, allowing the formulation of disposal measures based on specific warning information and avoiding blind disposal operations. Compared with traditional disposal solutions that rely on manual experience, disposal operations based on warning result data are more targeted and can adopt appropriate disposal measures based on different warning levels and risk types, such as adjusting storage environment parameters, transferring risky waste, and activating fire extinguishing equipment, thereby improving the effectiveness of disposal operations. The feedback optimization module optimizes the combustion risk prediction model based on the results of early warning and disposal operations, forming a closed-loop mechanism of "monitoring-early warning-disposal-optimization." Early warning results reflect the model's predictive accuracy in actual applications. If warning deviations occur, the cause of the deviation can be analyzed in conjunction with the disposal results, leading to model adjustments. Disposal results provide information correlating the actual development of risks with the effectiveness of disposal, providing practical data support for model optimization. Through this continuous feedback optimization, the model can continuously revise its predictive logic, improve its adaptability to complex risk situations, and ensure the accuracy of risk predictions in long-term applications. The parameter optimization module configures risk optimization goals based on historical warning data and historical disposal data and performs parameter-oriented optimization, which can achieve dynamic adjustment of system control parameters. Historical warning data contains risk warning records in different periods and environments, and historical disposal data records the application effects of different disposal measures. Through the analysis of these data, it is possible to find deficiencies in the current system parameter settings, such as unreasonable warning thresholds and disposal process parameters that need to be adjusted. Parameter-oriented optimization can accurately adjust system parameters according to risk optimization goals, such as reducing false alarm rates and improving disposal efficiency. This allows the system to continuously optimize its own operating parameters as actual application scenarios change and data accumulates, always maintain good risk control effects, and adapt to changes in control needs brought about by dynamic factors such as the storage environment and composition changes of battery production waste, ensuring the long-term stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1This is a timing diagram of the intelligent early warning and disposal system for battery production waste combustion risks according to the present invention; Figure 2 This is the flowchart of the early warning management module.

[0017] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0018] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0019] like Figure 1 As shown, the present application provides an intelligent early warning and disposal system for battery production waste combustion risks, including: a data acquisition module, a risk prediction module, an early warning management module, a disposal management module, a feedback optimization module and a parameter optimization module. The data acquisition module is used to obtain multi-source monitoring data of battery production waste within a first preset time period. The multi-source monitoring data includes physical and chemical parameters such as temperature, humidity, gas concentration and material accumulation state. These data are collected in real time through a sensor network deployed in the waste storage area and stored in the system database in the form of a time series; the risk prediction module trains a combustion risk prediction model based on the multi-source monitoring data and historical combustion event data. The model uses machine learning to A method is constructed and high-precision risk prediction is achieved through iterative optimization. The early warning management module obtains real-time monitoring data and performs early warning operations within a second preset time period, calculates the real-time risk value through the model, and generates an early warning signal by comparing it with the preset threshold. The disposal management module receives the early warning result data and performs disposal operations within a third preset time period, including initiating physical measures such as cooling, ventilation, or isolation. The feedback optimization module dynamically adjusts the parameters of the combustion risk prediction model according to the results of the early warning and disposal operations to improve the model's adaptability. The parameter optimization module configures risk optimization targets based on historical early warning data and historical disposal data, and adjusts the system operating parameters through an optimization algorithm to achieve overall performance improvement.

[0020] In one embodiment, Example 1: See Figure 2The data acquisition module deploys a multi-source sensor network in the battery production waste storage area. The network includes temperature sensors, humidity sensors, gas concentration detectors, and material status monitors. The sensors collect physical and chemical parameters at a fixed sampling frequency and form a time series data stream. The first preset time period is usually set as the historical data accumulation period, such as 30 consecutive days of monitoring data. The original data is transmitted to the central database for storage through a communication protocol. A data verification mechanism is used during the transmission process to prevent packet loss or misalignment. The data cleaning process performs noise filtering and outlier removal on the multi-source monitoring data. Noise data manifests as instantaneous pulse interference or sensor drift error. It is smoothed through a sliding window mean filter, and the window size is dynamically adjusted according to the sampling frequency. Outlier identification uses a threshold judgment method based on statistical distribution to calculate the standard deviation and mean of each parameter within the time window. Data points that exceed the range of three times the standard deviation are marked as outliers and replaced with the interpolation results of adjacent valid data. After cleaning, the monitoring data undergoes standardization to eliminate dimensional differences. Z-score normalization is used to ensure that the distribution of each parameter conforms to the zero-mean unit variance characteristic. Principal component analysis is used to map the data to feature vector space. Principal components with a cumulative contribution exceeding 85% are selected as features after dimensionality reduction to generate an initial feature vector of uniform dimension. A convolutional neural network receives this initial feature vector for deep feature extraction. The network structure consists of three convolutional layers and two max-pooling layers. The convolution kernel size is designed to be 3×3 and uses the ReLU activation function. The first convolutional layer outputs a feature map with compressed spatial dimensions through pooling. Subsequent convolutional layers gradually increase the number of channels to capture more complex nonlinear features. Finally, a fully connected layer flattens the high-dimensional feature map into a low-dimensional feature vector. This vector incorporates the spatiotemporal correlation characteristics of multi-source data and serves as the input to the risk prediction model.

[0021] The training process of the risk prediction module relies on paired samples of historical combustion event data and multi-source monitoring data. The historical combustion event data comes from the enterprise safety event database and includes a complete monitoring data sequence and event level labels for the 5 minutes before and after the combustion event. The combustion risk prediction model is constructed using a generative adversarial framework. The risk generator is a fully connected neural network structure. The number of nodes in the input layer matches the dimension of the feature vector. The hidden layer contains 128 neurons and uses dropout regularization to prevent overfitting. The output layer uses a sigmoid function to compress the predicted value to the range [0, 1] to represent the combustion probability. The risk discriminator adopts a convolutional neural network architecture. The input layer receives the output value of the risk generator and the corresponding historical combustion event label. After extracting discriminant features through two convolution operations, the fully connected layer outputs the true probability value, which reflects the degree of match between the generator's prediction and the actual event. The training phase employs an alternating optimization strategy. In the first iteration, the feature vector is fed into the risk generator to obtain a risk prediction value, while historical combustion event data is fed into the risk discriminator for binary classification. After the discriminator outputs the true probability value, the cross-entropy loss of the current batch prediction result is calculated, and the discriminator parameters are updated through backpropagation. In the second iteration, the discriminator parameters are fixed, and the generator's adversarial loss function is calculated based on the discriminant results. The optimization direction is to increase the probability that the risk prediction value output by the generator is judged by the discriminator to be a true event. A dynamic monitoring mechanism is set up for convergence determination. After every 10 iterations, a moving average of the true probabilities for 20 consecutive batches is calculated. Training is terminated when the absolute difference between this mean and the preset convergence threshold of 0.95 is less than or equal to the preset difference threshold of 0.02 for five consecutive iterations. After training, the generator possesses independent prediction capabilities, and the discriminator is used for subsequent model verification.

[0022] During the model deployment phase, a dual validation mechanism was established. The offline validation set used historical data not previously used in training. After inputting the feature vector, the mean squared error (MSE) between the risk generator output and the actual event label was compared. Online validation was performed via real-time data stream injection. When the true probability of the discriminator's response to the generator output remained consistently below 0.92, the model retraining process was triggered. Dynamic maintenance was implemented for the feature vector space mapping parameters, and the covariance matrix of the principal component analysis model was updated quarterly to ensure the adaptability of the feature distribution of newly collected data. The convolutional neural network weights were fine-tuned monthly through transfer learning, using the most recent three months of data as an incremental training set to maintain feature extraction capabilities. An expert review mechanism was introduced into the labeling process for historical combustion event data. Safety engineers manually verified automatically labeled event samples, corrected incorrect labels, and supplemented the event description text. The sample library was versioned, with each model iteration corresponding to an independent data snapshot to facilitate tracing the root cause of changes in predictive performance. The structural optimization of the risk discriminator adopts the channel attention mechanism, adding the SE module after the convolutional layer to automatically learn the feature channel weights, thereby improving the ability to capture key discriminant features; the loss function in the training process introduces FocalLoss to solve the imbalance problem of positive and negative samples and reduce the interference of high-frequency non-event samples on model optimization.

[0023] In one embodiment, Example 2: The early warning management module is activated and operates within a second preset time period, which is set as the system's real-time monitoring cycle. Sensor data streams from the current battery production waste storage area are typically acquired at a frequency of minutes. Real-time monitoring data is transmitted to the edge computing node via the industrial Internet of Things gateway. The data format is completely consistent with the historically collected multi-source monitoring data, including physical parameters such as temperature, volatile gas concentration, and infrared thermal imaging of the material surface. A risk generator is independently loaded from the combustion risk prediction model into the early warning subsystem memory. This generator uses the same neural network structure and weight parameters as the training phase to ensure consistency in prediction logic. The real-time monitoring data first passes through a normalization processing module, performing the same Z-score normalization operation as the training data to eliminate the impact of dimensional differences on the prediction. The normalized data is input into a feature extraction unit, which reuses the principal component analysis model and convolutional neural network parameters from the training phase to generate a real-time feature vector through the same dimensionality reduction and feature abstraction process. The feature vector is input into the risk generator, which outputs a combustion risk warning value, a floating-point number in the range [0, 1] whose value directly reflects the estimated probability of a combustion event occurring in the current waste pile.

[0024] The construction process of the rule mapping processing unit starts after the first warning operation. The system captures the current batch of real-time monitoring data and its corresponding combustion risk warning value. The preprocessing stage performs data cleaning operations on the real-time monitoring data. The cleaning rules are exactly the same as those in the training stage, including using sliding window filtering to eliminate high-frequency noise and eliminating abnormal sampling points based on the principle of three times the standard deviation. The feature extraction tool calls the deployed convolutional neural network model to convert the cleaned data into a set of feature fields. The dimension of the feature field is consistent with the dimension of the feature vector output in the training stage, including engineering features such as temperature gradient, gas concentration accumulation, and material porosity change rate. The combustion risk warning value is converted into a warning field through numerical discretization. The discretization strategy uses the equal-width binning method to divide the [0,1] interval into three warning fields: low risk (0-0.3), medium risk (0.3-0.7), and high risk (0.7-1.0). The mapping relationship is initialized using a random generation algorithm to establish an association matrix between the warning field and the feature field in memory. The matrix meets three constraints: each warning field is associated with at least one feature field, at most all feature fields, and any feature field can only be associated with the same warning field once. After the initial mapping scheme is generated, the core logic of the rule mapping processing unit is constructed. This unit is essentially a rule inference engine based on key-value pairs. When a feature field combination is input, the corresponding warning field is output by matching the mapping relationship.

[0025] The risk discriminator is separated from the combustion risk prediction model and loaded into the optimization subsystem. The discriminator maintains the network structure and parameters after training convergence. The reverse optimization process first constructs a test data set, randomly samples 100 groups of data samples from the real-time monitoring data stream and inputs them into the rule mapping processing unit. The unit outputs the rule combustion risk warning value according to the current mapping relationship, which is the discretized warning field code. At the same time, the original real-time monitoring data is input into the risk discriminator. The discriminator calculates the true probability value based on the feature vector and the rule combustion risk warning value. The true probability reflects the credibility of the warning value as a real event determined by the discriminator. The system calculates the arithmetic mean of the true probability of the 100 groups of samples, and uses the absolute difference between the average value and the preset convergence threshold of 0.95 as the scoring indicator of the current rule mapping processing unit. When the score exceeds the preset difference threshold of 0.02, the mapping relationship optimization process is initiated. The optimization algorithm uses stochastic gradient descent with momentum. The gradient calculation is based on the partial derivatives of the scoring function with respect to each element of the mapping matrix, and is decomposed layer by layer into each mapping relationship weight through the back-propagation chain rule. The weight update process introduces a learning rate decay mechanism, with the initial learning rate set to 0.1 and exponentially decaying after each iteration. The mapping relationship adjustment follows discrete constraints, and a constraint check is performed after each update: If the number of feature fields associated with a warning field is zero, an unoccupied feature field is randomly assigned to it. If a feature field is associated with multiple warning fields, only the optimal association in the gradient descent direction is retained. The optimization iteration continues until the moving average of three consecutive scoring rounds stabilizes within the difference threshold. The final solidified mapping relationship is written to the configuration memory of the rule mapping processing unit.

[0026] The rule mapping processing unit is deployed using a dual-mode operation mechanism, executing two computational paths in parallel during real-time warning operations: a primary path outputs precise combustion risk warning values ​​through the risk generator, while a secondary path outputs discretized warning fields through the rule mapping processing unit. The system compares the outputs of the two paths. When the warning field and the combustion risk warning value fall within the same risk interval, the warning is recorded as a valid rule mapping event. If there is a discrepancy between the intervals, the mapping relationship verification process is automatically triggered. The unit maintenance module implements periodic self-check tasks, loading the latest 24 hours of warning data at dawn each day and re-performing the scoring calculation and gradient descent optimization to ensure that the mapping relationship adapts to data distribution drift. Mapping scheme changes are version-controlled, generating incremental configuration files with each update and retaining the ability to roll back historical versions. The dynamic management mechanism for feature fields is activated monthly. When new monitoring indicators are added to the sensor network, the system automatically expands the feature field dimensions and initializes the mapping relationship for the newly added fields. If a feature field's coefficient of variation remains below 5% for 30 consecutive days, it is automatically disassociated from the warning field and marked as pending observation. The warning field division strategy supports online adjustment. When high-risk events occur frequently, the system automatically adjusts the lower limit of the high-risk interval from 0.7 to 0.6, regenerates the mapping plan, and performs an optimized verification process.

[0027] A lithium battery factory triggered an early warning process in a waste cathode material storage area. At 2:30 p.m., a temperature sensor detected a sudden increase in the local temperature from 42°C to 58°C within 10 minutes. Simultaneously, a volatile organic compound (VOC) concentration detector indicated that the ether concentration reached 1200 ppm. Real-time monitoring data packets, including timestamps, spatial coordinates, and 12 sensor readings, were transmitted to the early warning management module via an industrial IoT edge node. The risk generator loaded pretrained model parameters and normalized the input data: the temperature value was subtracted from the training set mean of 35°C and divided by the standard deviation of 8°C. Gas concentration values ​​were logarithmically transformed to eliminate dimensionality. This processed data was then fed into a feature extraction unit, which used stored convolutional neural network weights. After three layers of convolution, the unit output a feature vector of [0.24, -0.17, 0.83, 0.05, ...]. This feature vector was fed into the risk generator neural network, and after a fully connected layer, the output was a combustion risk warning value of 0.78. This value exceeded the high-risk threshold of 0.7, triggering a red alert.

[0028] The rule mapping processing unit construction program was immediately launched. The system extracted the current raw monitoring data: temperature 58°C, temperature rise rate 1.6°C / min, ether concentration 1200 ppm, material pile height 1.8 m, and other parameters. The feature extraction tool reused the same convolutional neural network model to convert the raw data into a set of feature fields [temperature gradient 0.84, concentration accumulation 0.92, bulk density 0.31, etc.]. The combustion risk warning value of 0.78 was discretized and converted into a warning field. The system's preset discretization rules defined the interval [0.7-1.0] as a high-risk code "H," generating the warning field "H-78" (H represents high risk, 78 represents the percentage of the original value). Mapping relationships were initialized using a random matrix algorithm, creating a 3×12 association matrix in memory (3 warning fields × 12 feature fields). The random assignment results are shown below: the high-risk field is associated with temperature gradient and concentration accumulation, the medium-risk field is associated with bulk density and porosity, and the low-risk field is associated with humidity change rate. This allocation satisfies the constraints that each warning field is associated with at least one feature field, is associated with at most all feature fields, and has no repeated associations.

[0029] The risk discriminator loads the validation model and scores the current mapping scheme. The system retrieves the most recent 50 sets of monitoring data from the real-time database and inputs them into the rule mapping processing unit. Based on the initial mapping relationship, the unit outputs a rule-based combustion risk warning value. For example, if a temperature gradient > 0.8 and a concentration accumulation > 0.9 are detected, an "H" code is output. The risk discriminator receives both the original feature vector and the rule warning code and calculates the true probability value. The discriminator outputs a value close to 1 when an actual combustion event occurs and close to 0 when no event occurs. The average true probability of the 50 data sets within the scoring calculation window is 0.89, with an absolute difference of 0.06 from the preset convergence threshold of 0.95. Exceeding the preset difference threshold of 0.02 triggers the optimization process.

[0030] Gradient descent optimization performed adjustments to the correlation matrix. The system calculated the partial derivatives of the score with respect to the matrix elements. It found that the weight associated with the temperature gradient and high-risk fields needed to be increased, while the weight associated with the bulk density and high-risk fields needed to be decreased. Weight updates employed a momentum acceleration mechanism. In the first iteration, the weight associated with the temperature gradient and high-risk fields increased from 0.37 to 0.42, and the weight associated with the bulk density and high-risk fields decreased from 0.21 to 0.16. Constraints were then verified to ensure that each warning field maintained at least one valid association. The optimization process continued for seven rounds, with the score gradually decreasing from 0.06 to 0.018. The final correlation matrix was adjusted to: the high-risk field was strongly associated with the temperature gradient (weight 0.51) and the concentration accumulation (0.49); the medium-risk field was associated with the bulk density (0.38) and porosity (0.33); and the low-risk field was associated with the humidity change rate (0.41) and other minor features.

[0031] When the new mapping scheme was put into operation, similar conditions were detected again at 3:10 PM that same day: a temperature gradient of 0.82 and a cumulative concentration of 0.88. Based on the optimized correlations, the rule-based mapping processing unit immediately output an "H" warning field. The risk discriminator verified that the true probability of this output was 0.94, and the score of 0.01 met the requirements. The system maintenance module performed a periodic self-check in the early morning of the same day, loading data from 32 warning events that had occurred in the previous 24 hours to recalculate the score. When the score rose to 0.025 in the evening, it automatically initiated a gradient descent fine-tuning, adjusting the temperature gradient weight from 0.51 to 0.53. The following week, when new infrared thermal imaging monitoring indicators were added, the feature field dimensions were expanded to 13. The system automatically initialized the weights associated with the newly added feature "hot spot area ratio" and each warning field, and performed a lightweight optimization verification to ensure that the added field did not affect overall mapping performance.

[0032] In one embodiment, Example 3: The scoring calculation process of the rule mapping processing unit is established on the basis of real-time data stream. The system extracts the latest 200 sets of real-time monitoring data and their corresponding rule combustion risk warning values ​​from the cache area of ​​the warning management module. Each set of data contains a complete set of feature fields and a discrete warning code generated by the rule mapping processing unit; the scoring algorithm adopts a sliding window mechanism, and the window size is set to 50 sets of data, sliding forward in steps of 10 sets to ensure that the scoring results reflect the recent system performance. The authenticity verification of the rule combustion risk warning value depends on the computing power of the risk discriminator. The risk discriminator loads the model parameters after training convergence. Its network structure includes three convolutional layers and two fully connected layers. The input data is a concatenated vector of feature fields and rule warning values; the discriminator outputs a true probability value between 0 and 1. The closer the value is to 1, the more consistent the rule warning value is with the characteristics of the real event. The scoring function is defined as the average value of the absolute difference between the true probability of all data points in the window and the preset convergence threshold. The calculation formula is:

[0033] in: Indicates the current score of the rule mapping processing unit, is the number of data samples in the sliding window (fixed at 50), Represents the risk discriminator for the The true probability value of the sample output is This is the system's preset convergence threshold (valued at 0.95). Scoring calculations are performed every 5 minutes. Mapping relationship optimization is triggered when three consecutive scoring results exceed the preset difference threshold of 0.02.

[0034] The gradient descent optimization process operates on the correlation matrix of the mapping scheme, which has a dimension of m×n (m is the number of warning fields, n is the number of feature fields), and the matrix elements Represents the association weight between the j-th warning field and the k-th feature field; the optimization goal is to minimize the partial derivative of the scoring function S with respect to the association matrix, and the gradient direction is calculated by the back propagation algorithm. Each iteration first calculates the score function for each weight The partial derivative of , the partial derivatives are calculated using numerical difference method, and the gradient is estimated by fine-tuning the weight value and observing the score change; the weight update formula is ,in is the learning rate (initial value is 0.1), is the momentum factor (value is 0.9), Represents the weight update from the previous iteration. After weight adjustment, constraint enforcement is performed: first, each warning field is associated with at least one feature field. For warning fields with zero associations, a randomly selected, partially occupied feature field is associated with the field. Second, each feature field is associated with only one warning field. For conflicting associations, the association with the largest absolute weight is retained. The iteration termination condition is set to either a score value less than or equal to 0.02 for five consecutive times or the maximum number of iterations, 100.

[0035] During the mapping relationship optimization process, a real-time monitoring mechanism is established. The system uses the current mapping scheme to process the test data set immediately after each weight update. The test data set contains 100 sets of fresh real-time monitoring data that have not participated in the optimization. The rapid evaluation module calculates the test score. When the test score drops by more than 10% compared to the pre-optimization level, the new mapping scheme is immediately applied. Otherwise, it rolls back to the previous version and reduces the learning rate to re-optimize. Historical score records are saved to the performance database to form score time series data. The trend analysis module uses a simple moving average method to predict the direction of score changes. When the prediction shows that the score will continue to deteriorate, the optimization process is started in advance. The mapping scheme version management adopts an incremental storage strategy. After each successful optimization, a difference configuration file is generated to record the association relationship with weight changes of more than 5%. The rollback mechanism can restore to any historical version within 10 seconds to ensure the system's rapid recovery capabilities in the event of optimization failure.

[0036] A dynamic learning rate adjustment strategy is implemented based on the rate of score change. When the score drops by more than 15% between two consecutive iterations, the learning rate is increased by 20% to accelerate convergence; when the score drops by less than 5%, the learning rate is reduced by 30% to improve stability. The momentum factor is dynamically adjusted based on the consistency of the weight update direction. When the weight update direction remains the same for three consecutive iterations, the momentum factor is increased by 0.1 to enhance inertia. If the update direction oscillates, the momentum factor is reset to 0.9. A soft constraint mechanism is introduced in the constraint enforcement phase, allowing temporary violations of the constraint that "each feature field is associated with only one warning field." However, a penalty term is added to the scoring function to measure the constraint violation. The penalty factor increases linearly with the number of iterations, ultimately forcing the solution to satisfy all constraints. After the optimization process, validation testing is performed to evaluate the performance of the new mapping solution using an independent validation set (consisting of 500 sets of recent real-time data). Validation metrics include score value, consistency ratio between warning fields and risk generator output, and processing latency. When the verification score exceeds the threshold, secondary optimization is initiated, using a more conservative learning rate (initial value is 0.01) and a smaller momentum factor (0.8). The final optimization results are written to the runtime memory of the rule mapping processing unit and the configuration file in the persistent storage is updated.

[0037] In one embodiment, Example 4: The disposal management module starts the classification process when receiving the warning result data, which includes the timestamp, the original reading of the sensor, the combustion risk warning value output by the risk generator, and the warning field code generated by the rule mapping processing unit; the classification judgment first checks the rule mapping processing unit score record corresponding to the real-time monitoring data, and the score is stored in the system performance database and the data collection time is marked. If the absolute difference between the score of the batch data and the preset convergence threshold of 0.95 is less than or equal to 0.02, the system marks the event as a candidate for a rule risk event; to verify the reliability of the candidate event, a feature sequence similarity analysis is performed: the real-time monitoring data is input into the feature extraction module after standardization and cleaning, and the module reuses the convolutional neural network parameters in the training phase and outputs a feature vector containing 12 feature fields such as temperature change rate, gas concentration accumulation, and hot spot area on the material surface; the warning result data is simultaneously subjected to feature extraction of the same process to ensure the consistency of the feature space.

[0038] The feature sequence is constructed using a dynamic sorting algorithm. The system determines the ranking order based on the variance contribution of each feature field in the combustion event in the past 30 days. The contribution calculation is based on the eigenvalue decomposition of the covariance matrix of the feature field and the event label. See Table 1 for an example of feature sequence construction for a certain event, in which the feature fields of the real-time monitoring data are arranged in descending order of contribution, and the warning result data adopts the same ranking rule.

[0039] Table 1: Example table of feature sequence construction

[0040] Similarity calculations use a sliding extraction window, with the window size set to 1 / 3 the length of the feature sequence (four fields in this example), and the window sliding step along the sequence is one field. Each window captures a subsequence as a convolution sequence unit, and the feature combination within the unit contains the numerical relationship between adjacent fields. Unit similarity is calculated using a weighted cosine similarity algorithm, with weights taken from the position weight column in Table 1. The calculation is performed by first converting the corresponding unit feature values ​​of the two sequences into unit vectors, then calculating the cosine of the angle between the vectors and multiplying it by the weight coefficient. The final similarity value is the arithmetic mean of the similarities of all window units. The preset similarity threshold is 0.85. When the similarity value calculated for the example data in Table 1 is 0.87, the system classifies the warning result data as a rule risk event. When a rule-based risk event triggers a rule mapping processing unit, the unit invokes a predefined set of action instructions based on the warning field code: If the warning field indicates a low risk code (0-0.3), the system activates basic ventilation for 15 minutes; if it indicates a medium risk code (0.3-0.7), it initiates high-pressure ventilation and sprays flame retardant mist for 20 minutes; and if it indicates a high risk code (0.7-1.0), it immediately isolates the waste area and activates the foam fire extinguishing system. Action instructions are transmitted to the PLC controller via the industrial bus, and the execution status is reported in real-time to the action log. If the similarity value falls below 0.85, it is classified as a non-rule risk event, and the system switches to a risk generator-led action mode. The risk generator receives the feature vectors of the warning result data, outputs a high-precision combustion probability value, and generates a action parameter matrix. The matrix dimensions include the action intensity coefficient, duration factor, and resource allocation priority. The action engine analyzes the matrix parameters and dynamically combines action measures. For example, if the probability value is 0.65, a customized action instruction is generated: "Ventilation intensity coefficient 0.7 + mist spray volume 120L / min + duration 25 minutes." A real-time learning mechanism is activated during the irregular event handling process. Handling instructions and execution effect data are stored in a special case library and extracted and analyzed weekly by the model optimization module. Classification results are verified using a double-blind review mechanism. 10% of classified events are randomly sampled daily, and the raw data is sent to the safety engineer console. Engineers manually classify and label them based on their operational experience, and the system automatically compares the manual labels with the automatic classification results. When the difference rate exceeds 5% for three consecutive days, calibration of the classification algorithm parameters is triggered. The feature ranking order is updated monthly. The system analyzes the distribution of feature fields for all combustion events that month, recalculates variance contributions, and adjusts the ranking weights. The sliding window size changes adaptively based on the length of the feature sequence. When the system adds new monitoring indicators, causing the number of feature fields to increase to 15, the window size is automatically adjusted to 5 fields. The handling instruction set supports online editing. Safety supervisors can modify ventilation equipment operating parameters or increase or decrease fire extinguishing material types through the management interface, and the modification records are stored in the version control system.

[0041] In one embodiment, Example 5: The feature sequence similarity calculation is initiated in the classification process of the disposal management module. The system extracts a feature field set of real-time monitoring data and early warning result data from the cache area. Each set of data contains engineering feature values ​​of 12 dimensions. The sliding extraction window size is set to one-third of the feature sequence length. When the number of feature fields is 12, the window size is fixed to 4 fields, and the window slides along the feature sequence with a step size of 1 field. The convolution sequence conversion process divides the original feature sequence into multiple local units, such as unit A containing fields 1 to 4, unit B containing fields 2 to 5, and unit H containing fields 9 to 12, generating a total of 9 feature combination units with the same dimensions. The feature values ​​in each unit are normalized according to the position weight to eliminate the influence of different dimensions on the similarity calculation.

[0042] Unit similarity is calculated using a vector space projection method. The corresponding unit eigenvalues ​​of the two sequences are treated as n-dimensional vectors (n is the window size). The cosine of the angle between the vectors is calculated as the basic similarity. For unit A in a particular event, the eigenvector of the real-time monitoring data is [4.8, 1250, 15.3, -2.1], and the eigenvector of the warning result data is [4.7, 1320, 15.1, -1.9]. The vectors are first normalized using the L2 norm to eliminate amplitude differences. The quotient of the dot product divided by the modulus product is then calculated. The final unit similarity is factored in by the ranking weight. The weight coefficient is derived from the variance contribution of the feature field. The weight for unit A is the arithmetic mean of the weights of the four fields within the window: (0.23 + 0.19 + 0.17 + 0.11) / 4 = 0.175. The weighted similarity is the basic cosine similarity multiplied by the weight coefficient. The overall similarity is the arithmetic mean of the weighted similarities of all units. When the calculation of 9 units is completed, the system automatically outputs a similarity quantization value between 0 and 1.

[0043] The feedback optimization module is activated every six hours, collecting the results of both early warning and disposal operations within the given cycle. The early warning result dataset includes 12 indicators, including the number of warning triggers, the distribution of risk generator output values, and the frequency of rule mapping processing unit invocations. The disposal result dataset records nine parameters, including the type of disposal instruction, the response time of the executing device, and the change in risk value after disposal. The data fusion unit converts these two types of results into a feedback information matrix recognizable by the model. The rows of the matrix correspond to time windows, and the columns contain composite features such as the warning accuracy flag (1 when the actual combustion event matches a high-risk warning), the disposal delay time (in seconds), and the risk reduction slope. The optimization interface of the combustion risk prediction model receives the feedback information matrix and adjusts the model parameters through an incremental learning mechanism. This incremental learning uses a mini-batch gradient descent method, injecting 50 sets of feedback data at a time, and the learning rate is set to one-tenth of the standard training phase.

[0044] The parameter optimization module initiates risk optimization target configuration on the first day of each month. The system retrieves historical warning and disposal data from the previous month to construct a multi-objective optimization function. This objective function comprises three core dimensions: minimizing the missed alarm rate for high-risk events (the proportion of actual combustions without warnings), minimizing the false alarm rate for low-risk events (the proportion of events with no risk but triggering warnings), and maximizing the disposal success rate (the proportion of events with a risk reduction of more than 50% after disposal). Parameter-directed optimization utilizes an improved genetic algorithm (GA). The population size is initialized to 100 individuals, each encoding a set of training parameters for the risk prediction module. These parameters include six adjustable parameters: the learning rate range (0.001-0.1), the number of convolution kernels (16-64), and the number of neurons in the fully connected layers (64-256). The GA iterative process consists of three phases: selection, crossover, and mutation. The selection operation is based on individual fitness ranking, and the fitness function is the weighted harmonic mean of the three optimization objectives. The crossover operation uses a two-point crossover method, randomly selecting the encoding position of two parent individuals to exchange segments. The mutation operation randomly modifies a parameter value with a 5% probability. In each iteration, the 10 individuals with the highest fitness are retained and directly advanced to the next generation, while the remaining 90 individuals are generated through crossover selection. The optimization process terminates when the change in the optimal fitness for 20 consecutive generations is less than 0.1%. The resulting optimal parameter combination is automatically updated to the training configuration of the risk prediction module, triggering a lightweight retraining of the model: two rounds of training iterations are performed using the last three months of data, allowing for rapid adaptation of the new parameters while maintaining prediction accuracy.

[0045] The dynamic maintenance mechanism for the feature sequence is performed weekly. The system analyzes the distribution of feature fields for newly added combustion events and recalculates the variance contribution of each field. If a field's contribution falls below 2% for four consecutive weeks, the system automatically removes it from the feature sequence, freeing up the corresponding storage space. The sliding window resize strategy is linked to the number of feature fields. When the feature sequence length changes due to the addition or removal of fields, the window size is automatically rounded to one-third of the new length. The window sliding step is always maintained at one field to ensure the continuity of the feature combination. The dimensions of the feedback information matrix are dynamically expanded. When a new carbon dioxide fire extinguishing device is added to the disposal system, columns such as "CO2 injection volume" and "time to reach concentration" are added to the matrix. The historical data warehouse uses columnar storage to support the rapid addition of new indicator fields. The parameter optimization module establishes a traceability system for the optimization process. Each genetic algorithm iteration records data such as population fitness distribution, optimal individual parameter encoding, and computation time. Engineers can use three-dimensional scatter plots to visualize the impact of different parameter combinations on the optimization target. An exception circuit breaker mechanism is implemented during the model retraining phase. If incremental training causes the validation set accuracy to drop by more than 3%, the model automatically rolls back to the pre-optimization parameter version and issues a maintenance alert. The feature field weight coefficient is reviewed by the Security Expert Committee every quarter. The committee manually adjusts the weight distribution plan through the management interface. The system automatically synchronizes to the similarity calculation process and performs a 72-hour observation period test.

[0046] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An intelligent early warning and disposal system for battery production waste combustion risk, characterized by: include: a data acquisition module for acquiring multi-source monitoring data of battery production waste within a first preset time period; Risk prediction module, used to train combustion risk prediction models based on multi-source monitoring data; An early warning management module is used to obtain real-time monitoring data within a second preset time period and perform early warning operations on the real-time monitoring data; a disposal management module, configured to obtain warning result data within a third preset time period and perform disposal operations on the warning result data; Feedback optimization module, used to optimize the combustion risk prediction model based on the results of early warning operations and disposal operations; The parameter optimization module is used to configure risk optimization objectives based on historical warning data and historical disposal data and perform parameter-oriented optimization.

2. The intelligent early warning and disposal system for battery production waste combustion risk according to claim 1 is characterized in that: The data acquisition module maps the multi-source monitoring data to the feature vector space after preprocessing to obtain the feature vector; Preprocessing includes data cleaning to remove noise data and outliers from multi-source monitoring data to obtain cleaned monitoring data; Mapping the cleaned monitoring data to the feature vector space to obtain the initial feature vector; The initial feature vector is convolved based on the convolutional neural network to obtain the feature vector.

3. The intelligent early warning and disposal system for battery production waste combustion risk according to claim 2 is characterized in that: The risk prediction module is trained to obtain a combustion risk prediction model based on multi-source monitoring data and historical combustion event data; The combustion risk prediction model includes a risk generator and a risk discriminator; The feature vector is input into the risk generator to obtain the risk prediction value, and the historical combustion event data is input into the risk discriminator to make a true probability judgment based on the risk prediction value; If the absolute difference between the mean of the true probability and the preset convergence threshold for multiple consecutive times is less than or equal to the preset difference threshold, the training converges and the combustion risk prediction model is obtained.

4. The intelligent early warning and disposal system for battery production waste combustion risk according to claim 3 is characterized in that: The early warning management module extracts the risk generator and risk discriminator of the combustion risk prediction model; Based on the risk generator, early warning is performed on the real-time monitoring data to obtain the combustion risk warning value; Based on real-time monitoring data and combustion risk warning values, a rule mapping processing unit is constructed; The mapping relationship of the processing unit is mapped based on the reverse optimization rule of the risk discriminator.

5. The intelligent early warning and disposal system for battery production waste combustion risk according to claim 4 is characterized in that: Based on real-time monitoring data and combustion risk warning values, a rule mapping processing unit is constructed, including: After preprocessing the real-time monitoring data, feature extraction is performed on the preprocessed data using a feature extraction tool to obtain multiple feature fields; Extract multiple warning fields from the combustion risk warning value; Establish a mapping relationship between feature fields and warning fields, including: Initialize and generate a mapping scheme that meets the constraint conditions; wherein the mapping scheme includes a mapping relationship between each warning field and a number of feature fields; Constraints include that each warning field must have a mapping relationship with at least one feature field, each warning field must have a mapping relationship with at most all feature fields, and each warning field must have and can only have one mapping relationship with each feature field. A rule mapping processing unit is constructed based on the mapping scheme.

6. The intelligent early warning and disposal system for battery production waste combustion risk according to claim 5 is characterized in that: Reversely optimizing the mapping relationship of the rule mapping processing unit based on the risk discriminator, including determining the score of the rule mapping processing unit; Based on the rule mapping processing unit, a corresponding rule combustion risk warning value is generated for the real-time monitoring data; The risk discriminator based on the combustion risk prediction model determines the true probability of the rule combustion risk warning value; The absolute difference between the true probability and the preset convergence threshold is used as the score of the rule mapping processing unit; If the absolute difference between the score and the preset convergence threshold is greater than the preset difference threshold, the mapping relationship of the mapping scheme is updated based on the gradient descent method until the absolute difference between the score and the preset convergence threshold is less than or equal to the preset difference threshold.

7. The intelligent early warning and disposal system for battery production waste combustion risk according to claim 6 is characterized in that: The disposal management module classifies the warning result data into regular risk events and non-regular risk events; If it is a rule risk event, the warning result data is processed based on the rule mapping processing unit; If it is a non-regular risk event, the warning result data will be processed based on the risk generator.

8. The intelligent early warning and disposal system for battery production waste combustion risk according to claim 7 is characterized in that: The warning result data is classified into regular risk events and non-regular risk events, including: If the absolute difference between the score of the rule mapping processing unit and the preset convergence threshold is less than or equal to the preset difference threshold, the corresponding real-time monitoring data is regarded as a rule risk event; Both real-time monitoring data and early warning result data are pre-processed and feature extracted; Based on the order of the characteristic fields in the real-time monitoring data and the early warning result data, the corresponding characteristic sequences are constructed respectively; Calculate the similarity value of the corresponding feature sequence; If the similarity value is greater than or equal to the preset similarity threshold, the warning result data is classified as a regular risk event; otherwise, the warning result data is classified as a non-regular risk event.

9. The intelligent early warning and disposal system for battery production waste combustion risk according to claim 8, characterized in that: Calculate the similarity value of the corresponding feature sequence, including: Load the sliding extraction window to convert the feature sequence into a convolution sequence with a fixed dimension; Calculate the similarity value of the feature combination of each unit in the corresponding convolution sequence respectively; wherein the feature combination is extracted from the feature sequence based on the sliding extraction window; The similarity value is determined based on the average of the similarity values ​​of the feature combinations of all units.

10. The intelligent early warning and disposal system for battery production waste combustion risk according to claim 9, characterized in that: The feedback optimization module inputs the warning operation results and the disposal operation results as feedback information into the combustion risk prediction model to dynamically optimize the combustion risk prediction model; The parameter optimization module configures the risk optimization target based on the feedback information and performs parameter-oriented optimization. The parameter-oriented optimization result is used to update the training parameters of the risk prediction module.

Images