Intelligent monitoring method and system based on industrial solid waste data linkage and storage medium

Abstract

The application provides an intelligent monitoring method and system based on industrial solid waste data linkage and a storage medium. The monitoring method comprises the following steps: acquiring industrial solid waste multi-source heterogeneous time sequence data, and extracting data containing current and historical time as input time sequence features; by improving a light-weight convolution gate unit, the features are divided into three groups according to channels, forward, expansion and reverse deep convolution are performed respectively, wherein the expansion rate is determined according to the hidden state of the previous time, then point-by-point convolution fusion is performed, the fusion features and the hidden state of the previous time are combined to calculate the gate and the candidate hidden state, a loss function is constructed based on the conservation of solid waste mass, a modulation coefficient is adjusted to update the gate, and then the hidden state of the current time is obtained; the hidden state sequence is input into a full connection network, and an industrial solid waste state monitoring and early warning result is output.

Description

Technical Field

[0001] This application belongs to the field of monitoring, and in particular relates to an intelligent monitoring method, system and storage medium based on industrial solid waste data linkage. Background Technology

[0002] Industrial solid waste involves multiple stages and complex processes from generation, collection, storage, transportation to disposal. Monitoring systems generate massive amounts of multi-source, heterogeneous data, such as production data, transportation trajectories, inventory levels, and operating parameters of disposal facilities. Traditional monitoring methods often rely on human experience or threshold alarms, which cannot reliably handle such complex data and frequently suffer from problems such as data silos and information lag. Purely data-driven analysis models may produce results that contradict basic laws when making state predictions or detecting anomalies, leading to unreliable predictions and high false alarm rates, seriously affecting the scientific nature and accuracy of regulatory decisions.

[0003] Recurrent neural networks (RNNs) and their variants, such as gated recurrent units (GRUs), are applied to industrial time-series data analysis to detect dependencies in the data. Some studies further combine convolutional neural networks (CNNs) with RRNs, utilizing convolutional operations to extract local temporal features and reducing model complexity through lightweight design. Furthermore, to improve model interpretability and predictive reliability, the idea of ​​information neural networks can be adopted, incorporating a law-based residual penalty term into the model training loss function to guide the model to learn a mapping relationship that conforms to constraints. Existing convolutional recurrent structures mostly use fixed, single-scale convolutional kernels to extract temporal features, failing to detect multi-scale dependency patterns (long, medium, and short-term) based on changes in the input sequence. Existing models also employ a relatively coarse approach to pattern fusion, typically using it as a static global constraint during model training. This fails to achieve real-time consistency verification and modulation of the model's internal state at each moment during inference, leading to potential deviations in the internal information flow updates when facing data noise or sudden changes in operating conditions, thus limiting further improvements in model prediction accuracy and stability. Summary of the Invention

[0004] To address the problem that existing technologies cannot detect multi-scale dependency patterns of long, medium, and short term based on changes in the input sequence, and that the updates of internal information flow may deviate from reality when faced with data noise or sudden changes in operating conditions.

[0005] In the first aspect, this disclosure provides an intelligent monitoring method based on industrial solid waste data linkage, comprising the following steps: Acquire multi-source heterogeneous time-series data of industrial solid waste, and for each moment, extract the time-series data including the current moment and historical moments of a preset length as the input time-series feature for the current moment; The input temporal features are processed by an improved lightweight convolutional gating unit. The processing divides the input temporal features into a first feature subgroup, a second feature subgroup, and a third feature subgroup along the channel dimension. A temporal forward depthwise convolution is performed on the first feature subgroup, an expanded depthwise convolution is performed on the second feature subgroup, and a temporal backward depthwise convolution is performed on the third feature subgroup. The processing results of the three subgroups are fused using pointwise convolution to generate a fused feature. Combining the fusion features with the previous hidden state, the original update gate, reset gate, and candidate hidden state are calculated; based on the loss function derived from the preset solid waste mass conservation constraint, the value of the loss function corresponding to the candidate hidden state is calculated; according to the value of the loss function, a consistency modulation coefficient is generated through a preset monotonically decreasing function, and the modulation coefficient is multiplied element-wise with the original update gate to obtain the update gate; using the update gate and reset gate, combined with the previous hidden state and the candidate hidden state, the output hidden state at the current moment is generated; The output hidden state sequence is input into a fully connected network to obtain the monitoring and early warning results of the industrial solid waste status.

[0006] For example, dividing the input temporal features into a first feature subgroup, a second feature subgroup, and a third feature subgroup along the channel dimension includes: If the channel dimension of the input time series feature is C, then the first C divided by 3 and rounded down is assigned to the first feature subgroup, the middle C divided by 3 and rounded down is assigned to the second feature subgroup, and the remaining channels are assigned to the third feature subgroup.

[0007] For example, the expansion rate for the second feature subgroup is determined from a preset expansion rate set based on the channel subgroup activation energy distribution of the hidden state at the previous time step, including: Divide the hidden state of the previous time step into K non-overlapping channel subgroups along the channel dimension; Calculate the L2 norm of the eigenvalues ​​within each channel subgroup as the activation energy of the subgroup; Select the index k corresponding to the channel subgroup with the highest activation energy; Select the kth expansion rate from the preset expansion rate set as the expansion rate for the second feature subgroup at the current time.

[0008] For example, the step of generating fused features by fusing the results of three subgroups using pointwise convolution includes: Apply a depthwise convolution with a kernel size of 3×1 to the first feature subgroup; Apply dilated depthwise convolution with a dilation rate and a kernel size of 3×1 to the second feature subgroup; The third feature subgroup is flipped in the time dimension, and then a depthwise convolution with a kernel size of 3×1 is applied to the flipped third feature subgroup. The result is then flipped again in the time dimension to restore the original order. The outputs of the three depthwise convolutions are concatenated along the channel dimension, and then information is fused through a 1×1 pointwise convolution to generate the fused feature.

[0009] For example, calculating the value of the loss function corresponding to the candidate hidden state based on the loss function derived from the preset solid waste mass conservation constraint includes: The candidate hidden states are mapped to the predicted change in solid waste mass through an auxiliary fully connected layer; The loss function is the square of the difference between the predicted change in solid waste mass and the actual change in solid waste mass. Calculate the value of the loss function.

[0010] For example, generating consistency modulation coefficients based on the value of the loss function using a preset monotonically decreasing function includes: The uniformity modulation coefficient is obtained through the formula The calculation yields the result, where k is a hyperparameter. This represents the loss function value.

[0011] In the second aspect, this disclosure provides an intelligent monitoring system based on industrial solid waste data linkage, including the following modules: The extraction module is used to acquire multi-source heterogeneous time-series data of industrial solid waste, and for each moment, extract the time-series data including the current moment and historical moments of a preset length as the input time-series feature of the current moment; The execution module is used to process the input temporal features by improving the lightweight convolutional gating unit. The processing divides the input temporal features into a first feature subgroup, a second feature subgroup, and a third feature subgroup along the channel dimension; performs temporal forward depth convolution on the first feature subgroup, performs dilated depth convolution on the second feature subgroup, and performs temporal backward depth convolution on the third feature subgroup; and fuses the processing results of the three subgroups using pointwise convolution to generate fused features. The generation module is used to combine the fusion features with the hidden state at the previous time step to calculate the original update gate, reset gate, and candidate hidden state; calculate the value of the loss function corresponding to the candidate hidden state based on the loss function derived from the preset solid waste mass conservation constraint; generate a consistency modulation coefficient through a preset monotonically decreasing function according to the value of the loss function, and multiply the modulation coefficient with the original update gate element by element to obtain the update gate; and generate the output hidden state at the current time step by combining the update gate and reset gate with the hidden state at the previous time step and the candidate hidden state. The input module is used to input the output hidden state sequence into a fully connected network to obtain the monitoring and early warning results of the industrial solid waste status.

[0012] Further, the step of dividing the input temporal features into a first feature subgroup, a second feature subgroup, and a third feature subgroup along the channel dimension includes: If the channel dimension of the input time series feature is C, then the first C divided by 3 and rounded down is assigned to the first feature subgroup, the middle C divided by 3 and rounded down is assigned to the second feature subgroup, and the remaining channels are assigned to the third feature subgroup.

[0013] Furthermore, the expansion rate for the second feature subgroup is determined from a preset expansion rate set based on the channel subgroup activation energy distribution of the hidden state at the previous time step, including: Divide the hidden state of the previous time step into K non-overlapping channel subgroups along the channel dimension; Calculate the L2 norm of the eigenvalues ​​within each channel subgroup as the activation energy of the subgroup; Select the index k corresponding to the channel subgroup with the highest activation energy; Select the kth expansion rate from the preset expansion rate set as the expansion rate for the second feature subgroup at the current time.

[0014] Furthermore, the process of fusing the results of the three subgroups using pointwise convolution to generate fused features includes: Apply a depthwise convolution with a kernel size of 3×1 to the first feature subgroup; Apply dilated depthwise convolution with a dilation rate and a kernel size of 3×1 to the second feature subgroup; The third feature subgroup is flipped in the time dimension, and then a depthwise convolution with a kernel size of 3×1 is applied to the flipped third feature subgroup. The result is then flipped again in the time dimension to restore the original order. The outputs of the three depthwise convolutions are concatenated along the channel dimension, and then information is fused through a 1×1 pointwise convolution to generate the fused feature.

[0015] Further, the calculation of the loss function value corresponding to the candidate hidden state based on the loss function derived from the preset solid waste mass conservation constraint includes: The candidate hidden states are mapped to the predicted change in solid waste mass through an auxiliary fully connected layer; The loss function is the square of the difference between the predicted change in solid waste mass and the actual change in solid waste mass. Calculate the value of the loss function.

[0016] Further, the step of generating consistency modulation coefficients based on the value of the loss function using a preset monotonically decreasing function includes: The uniformity modulation coefficient is obtained through the formula The calculation yields the result, where k is a hyperparameter. This represents the loss function value.

[0017] This invention proposes a data-driven monitoring method for industrial solid waste. The improved lightweight convolutional gating unit structure is simplified, has low computational overhead, and high operating efficiency. By dividing the input features into three subgroups and performing temporally forward, dilated, and temporally backward deep convolutions respectively, it can detect short-term, long-term, and contextual temporal dependencies in the data in parallel from different perspectives, achieving a deeper exploration of temporal features and improving the model's ability to represent complex temporal patterns. Furthermore, by utilizing the constraint of solid waste mass conservation to modulate the update gate, it ensures that the evolution of the model's hidden states follows fundamental laws, making the monitoring and early warning results not only data-driven but also reasonable and credible, enhancing the model's generalization performance and reliability and accuracy in real-world industrial scenarios. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a parallel multi-scale network architecture; Figure 2 This is a schematic diagram of a constraint modulation mechanism. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In the first embodiment, the present invention proposes an intelligent monitoring method based on industrial solid waste data linkage, comprising the following steps: S1, acquire multi-source heterogeneous time series data of industrial solid waste, and for each moment, extract the time series data including the current moment and historical moments of a preset length as the input time series feature of the current moment; Data is collected in real time from programmable logic controllers (PLCs), distributed control systems (DCS), and manufacturing execution systems (MES) in the industrial field via OPCUA, MQTT, or Modbus-TCP protocols. Data sources include, but are not limited to, numerical data from solid waste inbound weight sensors, component category data from sorting line infrared spectrometers, numerical data from incinerator temperature sensors, and textual data from solid waste source and disposal plans recorded in the ERP system. The collected multi-source heterogeneous data undergoes preprocessing. Z-score normalization (subtracting the mean and dividing by the standard deviation) is used for all numerical data. For categorical data such as component categories, one-hot encoding is used to convert the data into multi-dimensional binary vectors. For textual data, text cleaning is performed to remove special symbols, redundant characters, and invalid information. Then, Chinese word segmentation and stop word removal are used for text preprocessing. TF-IDF feature extraction is used to convert the preprocessed text information into fixed-length numerical feature vectors. Missing data is filled using linear interpolation. Set the historical time length L. For the current time t, use the data sliding window operation to extract the data from time t-L+1 to time t. Reshape this multivariate time series segment into a two-dimensional tensor with the number of channels C × the sequence length L, and use it as the input time series feature for the current time.

[0021] S2, the input temporal features are processed by an improved lightweight convolutional gating unit. The processing divides the input temporal features into a first feature subgroup, a second feature subgroup, and a third feature subgroup along the channel dimension. Forward temporal depthwise convolution is performed on the first feature subgroup, dilated depthwise convolution is performed on the second feature subgroup, and inverse temporal depthwise convolution is performed on the third feature subgroup. The dilation rate for the second feature subgroup is determined from a preset dilation rate set based on the channel subgroup activation energy distribution of the hidden state at the previous time step. The processing results of the three subgroups are fused using pointwise convolution to generate a fused feature. Input temporal features of dimension C×L On the channel dimension C, the torch.split function is used to equally divide the channel into three feature subgroups, each with a dimension of (C / 3)×L. , , Regarding the hidden state of the previous moment. Similarly, the hidden state is divided into a set with a preset expansion rate along the channel dimension. k subgroups with the same number of elements k , , , For each subgroup Calculate the L2 norm as the activation energy Select the index corresponding to the subgroup with the largest energy value, and from the set Extract the expansion rate corresponding to the index. The expansion rate is used as the second feature subgroup. Different channels correspond to solid waste data from different stages. Activating channels with high energy represents the current dominant factor, thus selecting the corresponding receptive field. Depth-first convolution is implemented using `torch.nn.Conv1d` with the `groups` parameter set to the number of input channels C / 3, and applied to the first feature subgroup. The execution core size is Step size is 1, fill size is Standard time forward depthwise convolution for the second feature subgroup The execution core size is Step size is 1, fill size is Expansion rate Dilated depthwise convolution for the third feature subgroup First, the `torch.flip` function is used to flip the result along the time dimension (L dimension). Then, a depthwise convolution with the same parameters as the first subgroup is performed, and the result is flipped back again to achieve temporally inverse depthwise convolution. The outputs of the three depthwise convolutions are concatenated along the channel dimension using the `torch.cat` function. A pointwise convolution with a kernel size of 1×1 (`torch.nn.Conv1d` with `kernel_size=1`) is used. The concatenated features are then fused and interacted with, resulting in an output dimension of... ×L fusion features .

[0022] In an alternative embodiment, dividing the input temporal features into a first feature subgroup, a second feature subgroup, and a third feature subgroup along the channel dimension includes: If the channel dimension of the input time series feature is C, then the first C divided by 3 and rounded down is assigned to the first feature subgroup, the middle C divided by 3 and rounded down is assigned to the second feature subgroup, and the remaining channels are assigned to the third feature subgroup.

[0023] The shape of the input temporal feature tensor is determined, typically by the batch size, time step, and number of channels C. For example, a typical input tensor... The dimensions could be 64, 24, or 128, where the batch size is 64, the time step is 24 time steps, and the channel dimension C is 128.

[0024] Calculate the number of channels in each subgroup according to the partitioning rules. Number of channels in the first feature subgroup. for That is, 42. The number of channels in the second feature subgroup. Also That is, 42. The number of channels in the third feature subgroup. That is the remaining part, i.e. =44. The method of rounding down and padding ensures that when 3 is not divisible by C, all channels are uniquely assigned, and the first two subgroups are of equal size. At the implementation level, segmentation can be achieved through tensor slicing operations: the first feature subgroup... Depend on The second feature subgroup consists of channels 0 to 41. Composed of channels 42 to 83; the third feature subgroup It consists of channels 84 to 127. The resulting three subgroups... , , The data will be fed into different convolutional branches for processing. This partitioning strategy ensures a balanced allocation of computing resources and full utilization of feature information.

[0025] In some embodiments, the expansion rate for the second feature subgroup is determined from a preset expansion rate set based on the channel subgroup activation energy distribution of the hidden state at the previous time step, including: Divide the hidden state of the previous time step into K non-overlapping channel subgroups along the channel dimension; Calculate the L2 norm of the eigenvalues ​​within each channel subgroup as the activation energy of the subgroup; Select the index k corresponding to the channel subgroup with the highest activation energy; Select the kth expansion rate from the preset expansion rate set as the expansion rate for the second feature subgroup at the current time.

[0026] Pre-set a set of expansion rates This set defines all possible expansion rates. For example, to detect short-term, medium-term, and long-term dependencies, you can set... At this point, the number of subgroups K is 4. The values ​​in this set typically grow exponentially or in the form of a Fibonacci sequence, achieving logarithmic receptive field coverage. Assume the hidden state of the previous time step... It is a tensor with the batch size as one dimension and the hidden layer dimension as another dimension, such as 64 or 256.

[0027] Will The hidden layer dimension of 256 is divided into K equal 4 non-overlapping channel subgroups, each with a dimension of 64. Activation energy is calculated for each channel subgroup. Activation energy is represented by the L2 norm, which is calculated as the square root of the sum of squares of all elements within the subgroup. For example, for the first subgroup... Activate energy Calculate the activation energy of all four subgroups in sequence. , , , Suppose that for a given sample, the calculated energy values ​​are 12.7, 25.1, 18.5, and 9.3. Select the subgroup with the highest energy value; in this example, the second subgroup, with an energy of 25.1, has an index k of 1. Based on this index k=1, select from the pre-defined expansion rate set. Select the corresponding expansion rate, i.e. The expansion rate for the second feature subgroup at the current time step is 2. This mechanism allows the model to automatically adopt a moderate expansion rate when it considers the intermediate-term dependence to be more important, resulting in a higher activation level in the second subgroup.

[0028] In some embodiments, the process of generating fused features by fusing the results of pointwise convolution of the three subgroups includes: Apply a depthwise convolution with a kernel size of 3×1 to the first feature subgroup; Apply dilated depthwise convolution with a dilation rate and a kernel size of 3×1 to the second feature subgroup; The third feature subgroup is flipped in the time dimension, and then a depthwise convolution with a kernel size of 3×1 is applied to the flipped third feature subgroup. The result is then flipped again in the time dimension to restore the original order. The outputs of the three depthwise convolutions are concatenated along the channel dimension, and then information is fused through a 1×1 pointwise convolution to generate the fused feature.

[0029] The segmented feature subgroups are processed through three parallel, functionally differentiated convolutional branches to detect comprehensive temporal information. These three branches, together with the subsequent fusion layer, constitute a parallel multi-scale convolutional module, which is a neural network. The first feature subgroup... Second feature subgroup and the third feature subgroup Their dimensions are respectively , and Where B is the batch size and T is the time step. , , This represents the number of channels in each subgroup. The network contains three parallel convolutional branches and one fusion layer. The first branch processes the first feature subgroup. The first branch applies standard depthwise convolution, focusing on extracting local, continuous positive temporal dependencies. This convolution kernel has a size of 3×1 and an dilation rate of 1. The second branch processes the second feature subgroup. Apply dilated depthwise convolution, with dilation rate As determined by the preceding steps, this branch can detect longer-range, discontinuous temporal patterns in a skip-based manner, and the kernel size is also 3×1. The third branch processes the third feature subgroup. To detect inverse time dependencies, the implementation first... The convolution result is flipped again in the time dimension by applying a standard 3×1 depthwise convolution. To maintain the time dimension length, all depthwise convolution operations use padding to make the output length equal to the input length. The outputs of the three branches... , , While maintaining the same dimensions as the input subgroups, concatenate them along the channel dimensions to obtain a combined feature tensor. .right A 1×1 pointwise convolution is applied for fusion, which linearly combines the convolutions along the channel dimension to achieve the fusion and interaction of temporal information extracted from the three branches at different scales and in different directions. This generates fused features. The dimension is consistent with the dimension of the model's hidden state, such as... Figure 1 As shown.

[0030] S3, combining the fusion features and the hidden state of the previous time step, calculate the original update gate, reset gate, and candidate hidden state; based on the loss function derived from the preset solid waste mass conservation constraint, calculate the value of the loss function corresponding to the candidate hidden state; according to the value of the loss function, generate a consistency modulation coefficient through a preset monotonically decreasing function, and multiply the modulation coefficient with the original update gate element by element to obtain the update gate; using the update gate and reset gate, combined with the hidden state of the previous time step and the candidate hidden state, generate the output hidden state of the current time step; The gating signal and candidate states are calculated using a fully connected layer torch.nn.Linear and activation functions, and the original update gate is then used. Reset the door Candidate hidden state ,in Let be the sigmoid function, tanh be the hyperbolic tangent function, W and b be the learnable weights and biases, and [,] denote feature concatenation. This represents element-wise multiplication. It also represents the candidate hidden state. Input a pre-trained decoder network, which consists of two torch.nn.Linear layers, to map the hidden states back to the spatial representation, thus obtaining the estimated change in solid waste mass at the current time. According to the law of conservation of mass Calculate the actual change in mass. ,in , , The actual intake, discharge, and treatment quality of solid waste obtained from time-series data; the value of the loss function. It is obtained by calculating the mean square error, that is Through a monotonically decreasing function Calculate the uniformity modulation coefficients, where k is a hyperparameter used to control the descent rate of the function. Use the `torch.mul` function to multiply these modulation coefficients element-wise with the original update gate to obtain the update gate. According to the update rules of the gated loop unit, through the formula The output hidden state at the current moment is calculated.

[0031] In an alternative embodiment, calculating the value of the loss function corresponding to the candidate hidden state based on the loss function derived from a preset solid waste mass conservation constraint includes: The candidate hidden states are mapped to the predicted change in solid waste mass through an auxiliary fully connected layer; The loss function is the square of the difference between the predicted change in solid waste mass and the actual change in solid waste mass. Calculate the value of the loss function.

[0032] The law of conservation of mass is introduced into the model to evaluate the rationality of candidate hidden states. This is implemented through an auxiliary fully connected layer, which is an independent neural network that stores the candidate hidden states at the current time step. , dimension Where B is the batch size. This represents the hidden state dimension. The network is a fully connected layer, also known as a dense layer. It contains a weight matrix. and a bias vector The forward propagation process can be represented as follows: The parameters of this layer are trained end-to-end along with the rest of the model. One dimension is... The output tensor is the predicted change in solid waste mass. .

[0033] After obtaining the network output, the loss is defined according to the law of conservation of mass. Obtain the actual mass of solid waste entering, leaving, and being treated from the input data, and calculate the actual mass change. The formula for calculating the loss is: The aforementioned The value represents the current candidate hidden state. The smaller the value of the deviation between the information contained and the law of conservation of mass, the better the consistency. In an alternative embodiment, the loss function is the square of the difference between the predicted change in solid waste mass at the previous time step and the known actual change in mass at the previous time step.

[0034] In an alternative embodiment, generating the consistency modulation coefficients based on the value of the loss function using a preset monotonically decreasing function includes: The uniformity modulation coefficient is obtained through the formula The calculation yields the result, where k is a hyperparameter. This represents the loss function value.

[0035] The function ensures that when When the consistency approaches 0, i.e., when the consistency is high, Approaching 1; while when When the inconsistency increases, Approaching 0. A hyperparameter k is set, which is a constant greater than zero, and it controls the modulation coefficient. Regarding the loss The sensitivity of k. The value of k needs to be determined based on The typical numerical range is adjusted, and optimization can usually be achieved experimentally on the validation set. A typical optimal range is 0.1 to 1.0. Assuming k = 0.5, and based on the previous step... =1.3. Calculate the product. =0.65. Taking the negative exponent of this product yields the consistency modulation coefficient. It is approximately equal to 0.522. The calculated value is... The value will subsequently be updated with the original gate. Perform element-wise multiplication to obtain the update gate. This indicates that due to the inconsistency in candidate hidden states, the model's acceptance of new information will be weakened, thus prioritizing the retention of historical state information that better conforms to the established patterns. The constraint propagation process from the latent space state to the mapping modulation factor described above is as follows: Figure 2 As shown.

[0036] S4, input the output hidden state sequence into the fully connected network to obtain the monitoring and early warning results of the industrial solid waste status.

[0037] The output-hidden-state sequence generated by the model at all time steps The hidden state at the last moment in the game As a feature representation of the entire sequence; Input a multilayer perceptron (MLP) defined by torch.nn.Sequential, with the following network structure: an input dimension equal to The network consists of a fully connected layer whose output dimension is equal to the number of hidden nodes, followed by a ReLU activation function, a Dropout layer to prevent overfitting, and a fully connected layer whose output dimension equals the number of alert levels. The network's output layer selects an activation function based on the task type. For a three-class alert task (normal, watch out, alarm), the Softmax function is used to output the probability of each level. For a binary anomaly detection task, the Sigmoid function is used to output an anomaly probability value. When this probability value exceeds a preset threshold, such as 0.8, an alert is triggered.

[0038] The experimental conditions were uniformly set as follows: the hourly operation data of a large urban solid waste treatment center for two consecutive years was used as the dataset, and the dataset was divided into 70% as the training set, 15% as the validation set, and 15% as the test set according to the time sequence. The Adam optimizer was used, with a learning rate of 0.001, a batch size of 64, and 100 training epochs. The evaluation metrics used are mean absolute error (MAE) and root mean square error (RMSE). The baseline model is a standard GRU network.

[0039] Using the basic GRU model as a benchmark, the MAE is 5.12 and the RMSE is 7.89.

[0040] By integrating a parallel multi-scale convolution module on the base GRU but using a fixed dilation rate (i.e., removing the dilation rate determination step), the model performance is improved to MAE 4.35 and RMSE 6.51.

[0041] Building upon this, the addition of an expansion rate determination mechanism further improved model performance, reducing MAE to 4.18 and RMSE to 6.23. The complete scheme incorporating a consistency modulation mechanism achieved the best performance, with an MAE of 3.91 and an RMSE of 5.85, and the loss function value... The mean on the test set is 0.21, which is much lower than the 1.5 or higher when no constraints are applied.

[0042] Compared to the basic GRU, the use of parallel multi-scale convolutional modules reduces prediction error, demonstrating the reliability of feature extraction through parallel forward, long-range, and inverse branches. Compared to a fixed dilation rate, the dilation rate mechanism allows the model to adjust its receptive field to detect key temporal dependencies, leading to further performance improvements. The introduction of a consistency modulation mechanism, using the mass conservation law as a soft constraint, corrects the model's prediction behavior, enabling it to generate more reasonable predictions and achieving maximum optimization in MAE and RMSE metrics.

[0043] In a second embodiment, the present invention also provides an intelligent monitoring system based on industrial solid waste data linkage, comprising the following modules: The extraction module is used to acquire multi-source heterogeneous time-series data of industrial solid waste, and for each moment, extract the time-series data including the current moment and historical moments of a preset length as the input time-series feature of the current moment; The execution module is used to process the input temporal features by improving the lightweight convolutional gating unit. The processing divides the input temporal features into a first feature subgroup, a second feature subgroup, and a third feature subgroup along the channel dimension; performs temporal forward depth convolution on the first feature subgroup, performs dilated depth convolution on the second feature subgroup, and performs temporal backward depth convolution on the third feature subgroup; and fuses the processing results of the three subgroups using pointwise convolution to generate fused features. The generation module is used to combine the fusion features with the hidden state at the previous time step to calculate the original update gate, reset gate, and candidate hidden state; calculate the value of the loss function corresponding to the candidate hidden state based on the loss function derived from the preset solid waste mass conservation constraint; generate a consistency modulation coefficient through a preset monotonically decreasing function according to the value of the loss function, and multiply the modulation coefficient with the original update gate element by element to obtain the update gate; and generate the output hidden state at the current time step by combining the update gate and reset gate with the hidden state at the previous time step and the candidate hidden state. The input module is used to input the output hidden state sequence into a fully connected network to obtain the monitoring and early warning results of the industrial solid waste status.

[0044] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0045] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent monitoring method based on industrial solid waste data linkage, characterized in that, Includes the following steps: Acquire multi-source heterogeneous time-series data of industrial solid waste, and for each moment, extract the time-series data including the current moment and historical moments of a preset length as the input time-series feature for the current moment; The input temporal features are processed by an improved lightweight convolutional gating unit. The processing divides the input temporal features into a first feature subgroup, a second feature subgroup, and a third feature subgroup along the channel dimension. A temporal forward depthwise convolution is performed on the first feature subgroup, an expanded depthwise convolution is performed on the second feature subgroup, and a temporal backward depthwise convolution is performed on the third feature subgroup. The processing results of the three subgroups are fused using pointwise convolution to generate a fused feature. Combining the fusion features with the previous hidden state, the original update gate, reset gate, and candidate hidden state are calculated; based on the loss function derived from the preset solid waste mass conservation constraint, the value of the loss function corresponding to the candidate hidden state is calculated; according to the value of the loss function, a consistency modulation coefficient is generated through a preset monotonically decreasing function, and the modulation coefficient is multiplied element-wise with the original update gate to obtain the update gate; using the update gate and reset gate, combined with the previous hidden state and the candidate hidden state, the output hidden state at the current moment is generated; The output hidden state sequence is input into a fully connected network to obtain the monitoring and early warning results of the industrial solid waste status.

2. The method according to claim 1, characterized in that, The step of dividing the input temporal features into a first feature subgroup, a second feature subgroup, and a third feature subgroup along the channel dimension includes: If the channel dimension of the input time series feature is C, then the first C divided by 3 and rounded down is assigned to the first feature subgroup, the middle C divided by 3 and rounded down is assigned to the second feature subgroup, and the remaining channels are assigned to the third feature subgroup.

3. The method according to claim 1, characterized in that, The expansion rate used for the second feature subgroup is determined from a preset expansion rate set based on the channel subgroup activation energy distribution of the hidden state at the previous time step, including: Divide the hidden state of the previous time step into K non-overlapping channel subgroups along the channel dimension; Calculate the L2 norm of the eigenvalues ​​within each channel subgroup as the activation energy of the subgroup; Select the index k corresponding to the channel subgroup with the highest activation energy; Select the kth expansion rate from the preset expansion rate set as the expansion rate for the second feature subgroup at the current time.

4. The method according to claim 3, characterized in that, The process of fusing the results of the three subgroups using pointwise convolution to generate fused features includes: Apply a depthwise convolution with a kernel size of 3×1 to the first feature subgroup; Apply dilated depthwise convolution with a dilation rate and a kernel size of 3×1 to the second feature subgroup; The third feature subgroup is flipped in the time dimension, and then a depthwise convolution with a kernel size of 3×1 is applied to the flipped third feature subgroup. The result is then flipped again in the time dimension to restore the original order. The outputs of the three depthwise convolutions are concatenated along the channel dimension, and then information is fused through a 1×1 pointwise convolution to generate the fused feature.

5. The method according to claim 1, characterized in that, The loss function, based on a preset solid waste mass conservation constraint, calculates the value of the loss function corresponding to the candidate hidden state, including: The candidate hidden states are mapped to the predicted change in solid waste mass through an auxiliary fully connected layer; The loss function is the square of the difference between the predicted change in solid waste mass and the actual change in solid waste mass. Calculate the value of the loss function.

6. The method according to claim 1, characterized in that, The step of generating consistency modulation coefficients based on the value of the loss function using a preset monotonically decreasing function includes: The uniformity modulation coefficient is obtained through the formula The calculation yields the result, where k is a hyperparameter. This represents the loss function value.

7. An intelligent monitoring system based on industrial solid waste data linkage, characterized in that, Includes the following modules: The extraction module is used to acquire multi-source heterogeneous time-series data of industrial solid waste, and for each moment, extract the time-series data including the current moment and historical moments of a preset length as the input time-series feature of the current moment; The execution module is used to process the input temporal features by improving the lightweight convolutional gating unit. The processing divides the input temporal features into a first feature subgroup, a second feature subgroup, and a third feature subgroup along the channel dimension; performs temporal forward depth convolution on the first feature subgroup, performs dilated depth convolution on the second feature subgroup, and performs temporal backward depth convolution on the third feature subgroup; and fuses the processing results of the three subgroups using pointwise convolution to generate fused features. The generation module is used to combine the fusion features with the hidden state at the previous time step to calculate the original update gate, reset gate, and candidate hidden state; calculate the value of the loss function corresponding to the candidate hidden state based on the loss function derived from the preset solid waste mass conservation constraint; generate a consistency modulation coefficient through a preset monotonically decreasing function according to the value of the loss function, and multiply the modulation coefficient with the original update gate element by element to obtain the update gate; and generate the output hidden state at the current time step by combining the update gate and reset gate with the hidden state at the previous time step and the candidate hidden state. The input module is used to input the output hidden state sequence into a fully connected network to obtain the monitoring and early warning results of the industrial solid waste status.

8. The system according to claim 7, characterized in that, The step of dividing the input temporal features into a first feature subgroup, a second feature subgroup, and a third feature subgroup along the channel dimension includes: If the channel dimension of the input time series feature is C, then the first C divided by 3 and rounded down is assigned to the first feature subgroup, the middle C divided by 3 and rounded down is assigned to the second feature subgroup, and the remaining channels are assigned to the third feature subgroup.

9. The system according to claim 7, characterized in that, The expansion rate used for the second feature subgroup is determined from a preset expansion rate set based on the channel subgroup activation energy distribution of the hidden state at the previous time step, including: Divide the hidden state of the previous time step into K non-overlapping channel subgroups along the channel dimension; Calculate the L2 norm of the eigenvalues ​​within each channel subgroup as the activation energy of the subgroup; Select the index k corresponding to the channel subgroup with the highest activation energy; Select the kth expansion rate from the preset expansion rate set as the expansion rate for the second feature subgroup at the current time.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by a processor, implements the method as described in any one of claims 1-6.

Images