Safety monitoring method and system for trichloroisocyanuric acid wastewater treatment

Abstract

The invention discloses a safety monitoring method and system for trichloroisocyanuric acid wastewater treatment, and relates to the technical field of data analysis. The method comprises the following steps: converting a vibration signal into a two-dimensional time-frequency image, and taking the two-dimensional time-frequency image as input of a plurality of pre-trained state recognition models to obtain prediction probabilities of a plurality of state categories; according to a preset integrated weight, performing weighted summation on the prediction probability of the target state category to obtain a probability score of the target state category; and outputting the state category corresponding to the maximum probability score as a detection result. State recognition is carried out on the two-dimensional time-frequency image through the pre-training model, prediction accuracy is improved through weighted integration, and early fine faults can be effectively recognized. Compared with a traditional threshold value method, the method has the advantages that the fault recognition time point is advanced from the obvious damage stage to the non-obvious damage stage, early warning is facilitated, and the sudden equipment fault risk is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and particularly to a safety monitoring method and system for the treatment of trichloroisocyanuric acid wastewater. Background Art

[0002] As a highly efficient chlorine-containing disinfectant, the high-salt and high-organic-concentration wastewater generated during the production process of trichloroisocyanuric acid poses a serious threat to the ecological environment. To achieve the resource treatment of wastewater and energy conservation and emission reduction, the mechanical vapor recompression (MVR) process is widely used due to its advantages of high efficiency and energy conservation. This process compresses and heats the secondary steam through a compressor to realize the recycling of thermal energy. As the core equipment of the MVR system, the stable operation of the compressor directly determines the efficiency and energy consumption level of the entire wastewater treatment system. As a key component of the compressor, the performance decline of the bearing is extremely likely to cause abnormal equipment vibration, efficiency reduction, and even sudden shutdown, which not only interrupts the wastewater treatment process but also may lead to safety accidents such as the leakage of high-temperature and high-pressure steam.

[0003] Currently, most enterprises still rely on manual inspection and regular maintenance to monitor the operating status of MVR compressors, which have problems such as a long detection cycle. Some automated monitoring solutions use threshold warnings, such as the monitoring method based on temperature sensors. However, early subtle abnormalities do not cause obvious characteristics during machine operation, and such threshold warning methods lack sensitivity to early potential faults. Traditional monitoring methods have the problem of lagging response in detecting early fault hazards, which will cause the hazards to gradually intensify with the operation of the equipment, resulting in serious consequences, such as unplanned shutdown, interruption of the wastewater treatment process, increased equipment repair costs, and even safety hazards such as the leakage of high-temperature and high-pressure steam. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem of lagging response mentioned in the above background art, and to propose a safety monitoring method and system for the treatment of trichloroisocyanuric acid wastewater.

[0005] In the first aspect of the implementation of the present invention, a safety monitoring method for the treatment of trichloroisocyanuric acid wastewater is provided, which uses the mechanical vapor recompression process to treat wastewater. The method includes: Obtain the vibration signal of the compressor, and preprocess the vibration signal to obtain a target signal; Convert the target signal into a two-dimensional time-frequency image; Use the two-dimensional time-frequency image as the input of multiple pre-trained state recognition models to obtain the prediction probabilities of multiple state categories; According to the preset integration weights, weighted sum the prediction probabilities of the target state category to obtain the probability score of the target state category; the target state category is any one of the multiple state categories; Output the state category corresponding to the maximum probability score as the detection result.

[0006] Optionally, preprocess the vibration signal to obtain a target signal, including: Extract a first vibration signal with a preset length from the vibration signal; Perform variational mode decomposition on the first vibration signal to obtain multiple intrinsic mode components; Calculate the Pearson correlation coefficient between each intrinsic mode component and the first vibration signal; Select the intrinsic mode components with the absolute value of the correlation coefficient greater than a preset threshold for signal reconstruction to obtain the target signal.

[0007] Optionally, the structures of multiple state recognition models are the same; any one state recognition model includes: An initial convolution module for performing shallow feature extraction on the two-dimensional time-frequency image using a single convolutional layer and performing spatial compression to obtain a first feature map; A feature extraction module for performing multi-scale feature extraction on the first feature map using convolutional layers with different kernel sizes to obtain a second feature map; An attention enhancement module for enhancing the features of the second feature map using an attention mechanism to obtain a third feature map; A shaping module for converting the third feature map into a feature vector; A classification module for mapping the feature vector to a probability vector of a preset scale using a fully connected network.

[0008] Optionally, the feature extraction module includes two cascaded multi-scale dense blocks; the structures of the two multi-scale dense blocks are the same, and the operation process of any one multi-scale dense block includes:

[0009] where X in is the input of this multi-scale dense block; Conv represents the convolution operation, the subscripts 3×3, 5×5, 7×7, and 1×1 represent the kernel sizes of the convolution, and the superscript 1 represents the convolution stride; concat represents concatenation in the channel dimension; MaxPool represents max pooling, the subscript 2×2 represents the pooling window size, and the superscript 2 represents the pooling stride; X 11 、X 12 、X 13 、X 14 、X 21 、X 22 、X 23 and X 24 are the feature maps generated during the operation process, and X out is the output of this multi-scale dense block.

[0010] Optionally, the integrated weights are determined by the training effects of multiple state recognition models. Specifically, after the multiple state recognition models are trained, the particle swarm optimization algorithm is used to optimize and determine the integrated weights, including: Step 1: Use multiple state recognition models to predict each sample in the validation set to obtain the probability vector corresponding to each sample; Step 2: Initialize the population size, particle positions, and velocities; Step 3: Use the softmax function to convert the particle positions into effective weight distributions; Step 4: According to the effective weight distribution, integrate the probability vectors of each sample into probability scores, obtain the final prediction result, and calculate its accuracy as the fitness; Step 5: Update the individual optimal position and the global optimal position according to the fitness value; Step 6: Determine whether the preset stop condition is reached; if so, end the optimization process; otherwise, go to Step 7; Step 7: Update the particle positions according to the individual optimal position and the global optimal position, and return to Step 3.

[0011] In the second aspect of the implementation of the present invention, a safety monitoring system for trichloroisocyanuric acid wastewater treatment is provided, which is used to implement the above-mentioned safety monitoring method for trichloroisocyanuric acid wastewater treatment. It is characterized in that the system includes: A data acquisition module for acquiring the vibration signal of the compressor; A preprocessing module for preprocessing the vibration signal to obtain a target signal; A two-dimensional characterization module for converting the target signal into a two-dimensional time-frequency image; A base model decision module for using the two-dimensional time-frequency image as the input of multiple pre-trained state recognition models to obtain the prediction probabilities of multiple state categories; An integration module for weighted summing the prediction probabilities of the target state category according to the preset integrated weights to obtain the probability score of the target state category; the target state category is any one of the multiple state categories; A result output module for outputting the state category corresponding to the maximum probability score as the detection result.

[0012] Optionally, the preprocessing module includes: A segment extraction module for extracting a first vibration signal with a preset length from the vibration signal; A modal decomposition module for performing variational modal decomposition on the first vibration signal to obtain multiple intrinsic mode components; A correlation calculation module for calculating the Pearson correlation coefficient between each intrinsic mode component and the first vibration signal; A signal reconstruction module, configured to perform signal reconstruction on the intrinsic mode components with the absolute value of the correlation coefficient greater than a preset threshold to obtain a target signal.

[0013] Optionally, the structures of multiple state recognition models are the same; any one state recognition model includes: An initial convolution module, configured to perform shallow feature extraction on the two-dimensional time-frequency image by using a single convolutional layer and perform spatial compression to obtain a first feature map; A feature extraction module, configured to perform multi-scale feature extraction on the first feature map by using convolutional layers with different kernel sizes to obtain a second feature map; An attention enhancement module, configured to perform feature enhancement on the second feature map by using an attention mechanism to obtain a third feature map; A shaping module, configured to convert the third feature map into a feature vector; A classification module, configured to use a fully connected network to map the feature vector to a probability vector of a preset scale.

[0014] Optionally, the feature extraction module includes two cascaded multi-scale dense blocks; the structures of the two multi-scale dense blocks are the same, and the operation process of any one multi-scale dense block includes:

[0015] Wherein, X in is the input of this multi-scale dense block; Conv represents a convolution operation, the subscripts 3×3, 5×5, 7×7, and 1×1 represent the kernel sizes of the convolutional kernels, and the superscript 1 represents the convolution stride; concat represents concatenation in the channel dimension; MaxPool represents max pooling, the subscript 2×2 represents the pooling window size, and the superscript 2 represents the pooling stride; X 11 , X 12 , X 13 , X 14 , X 21 , X 22 , X 23 , and X 24 are the feature maps generated during the operation process, and X out is the output of this multi-scale dense block.

[0016] Optionally, the system further includes an integrated optimization module, configured to optimize and determine the integrated weights by using a particle swarm optimization algorithm after multiple state recognition models are trained; the operation process of the integrated optimization module includes: Step 1, use multiple state recognition models to predict each sample in the validation set to obtain a probability vector corresponding to each sample; Step 2, initialize the population size, particle positions, and velocities; Step 3: Use the softmax function to convert the particle position into an effective weight distribution; Step 4: According to the effective weight distribution, integrate the probability vectors of each sample into a probability score to obtain the final prediction result, and calculate its accuracy as the fitness; Step 5: Update the individual optimal position and the global optimal position according to the fitness value; Step 6: Determine whether the preset stop condition is reached; if so, end the optimization process; otherwise, go to Step 7; Step 7: Update the particle position according to the individual optimal position and the global optimal position, and return to Step 3.

[0017] Advantages of the present invention: The present invention proposes a safety monitoring method for trichloroisocyanuric acid wastewater treatment, which includes: acquiring the vibration signal of the compressor, preprocessing the vibration signal to obtain a target signal; converting the target signal into a two-dimensional time-frequency image; using the two-dimensional time-frequency image as the input of multiple pre-trained state recognition models to obtain the prediction probabilities of multiple state categories; according to the preset integration weights, weighted-sum the prediction probabilities of the target state category to obtain the probability score of the target state category; the target state category is any one of the multiple state categories; output the state category corresponding to the maximum probability score as the detection result.

[0018] By collecting the vibration signal of the compressor and converting it into a two-dimensional time-frequency image, using multiple pre-trained models for state recognition, and improving the prediction accuracy through weighted integration, early subtle faults can be effectively identified. Compared with the traditional threshold method, this method advances the fault recognition time point from the obvious damage stage to the non-obvious damage stage, which helps to give early warnings and reduce the risk of sudden equipment failures. Description of the Drawings

[0019] Figure 1 This is a flowchart of a safety monitoring method for trichloroisocyanuric acid wastewater treatment provided by an embodiment of the present invention; Figure 2 This is a network architecture diagram of a state recognition model provided by an embodiment of the present invention; Figure 3 This is a structural schematic diagram of a multi-scale dense block provided by an embodiment of the present invention; Figure 4 This is an architecture diagram of a safety monitoring system for trichloroisocyanuric acid wastewater treatment provided by an embodiment of the present invention. Detailed Embodiments

[0020]

[0020] To further illustrate the technical means and effects adopted by the present invention to achieve the intended invention purpose, the following will, in conjunction with the accompanying drawings and preferred embodiments, elaborate in detail on the specific implementation manners, structures, features and their effects according to the present invention as follows.

[0021] An embodiment of the present invention provides a safety monitoring method for the treatment of trichloroisocyanuric acid wastewater. Refer to Figure 1 , Figure 1 which is a flowchart of a safety monitoring method for the treatment of trichloroisocyanuric acid wastewater provided by an embodiment of the present invention. The method includes the following steps: S101, obtain the vibration signal of the compressor.

[0022] S102, perform preprocessing on the vibration signal to obtain a target signal.

[0023] S103, convert the target signal into a two-dimensional time-frequency image.

[0024] S104, take the two-dimensional time-frequency image as the input of multiple pre-trained state recognition models to obtain the prediction probabilities of multiple state categories.

[0025] S105, according to the preset integration weights, perform weighted summation on the prediction probabilities of the target state category to obtain the probability score of the target state category.

[0026] S106, output the state category corresponding to the maximum probability score as the detection result.

[0027] Among them, the target state category is any one of the multiple state categories.

[0028] Based on the safety monitoring method for the treatment of trichloroisocyanuric acid wastewater provided by the embodiment of the present invention, by collecting the vibration signal of the compressor and converting it into a two-dimensional time-frequency image, using multiple pre-trained models for state recognition, and improving the prediction accuracy through weighted integration, it can effectively identify early subtle faults. Compared with the traditional threshold method, this method advances the fault recognition time point from the stage with obvious damage to the stage with unobvious damage, which helps to give early warnings and reduce the risk of sudden equipment failures.

[0029] In one implementation, the continuous wavelet transform can be used to convert the target signal into a two-dimensional time-frequency image.

[0030] In one implementation, perform ensemble learning and probability weighting on the results of multiple state recognition models to improve the robustness and generalization ability of the overall judgment.

[0031] In one embodiment, step S102 includes: Step 1, extract a first vibration signal with a preset length from the vibration signal.

[0032] Step 2: Perform variational mode decomposition on the first vibration signal to obtain multiple intrinsic mode components.

[0033] Step 3: Calculate the Pearson correlation coefficient between each intrinsic mode component and the first vibration signal.

[0034] Step 4: Select the intrinsic mode components with the absolute value of the correlation coefficient greater than a preset threshold for signal reconstruction to obtain the target signal.

[0035] In this embodiment, by performing variational mode decomposition on the vibration signal and using the Pearson correlation coefficient to screen out the intrinsic mode components highly correlated with the original signal for reconstruction, noise and redundant information are effectively removed, key features are extracted, the quality of the target signal is improved, thereby enhancing the accuracy and robustness of subsequent state recognition.

[0036] In one implementation, the preset length is 1024 and the preset threshold is 0.2.

[0037] In one embodiment, the structures of multiple state recognition models are the same but the weights are different. Refer to Figure 2 , Figure 2 which is the network architecture diagram of a state recognition model provided by an embodiment of the present invention. As shown in the figure, any state recognition model includes an initial convolution module, a feature extraction module (the first multi-scale dense block + the second multi-scale dense block), an attention enhancement module, a shaping module, and a classification module, where: The initial convolution module is used to perform shallow feature extraction on the two-dimensional time-frequency image using one layer of convolutional layer and perform spatial compression to obtain the first feature map. Specifically, a 3×3 convolutional layer is used for shallow feature extraction, and a max pooling layer with a stride of 2 is used for spatial compression.

[0038] The feature extraction module is used to perform multi-scale feature extraction on the first feature map using convolutional layers with different kernel sizes to obtain the second feature map. Specifically, the feature extraction module includes two cascaded multi-scale dense blocks, and the two multi-scale dense blocks have the same structure but different weights. Refer to Figure 3 , Figure 3 which is the structural schematic diagram of a multi-scale dense block provided by an embodiment of the present invention. The operation process of any multi-scale dense block includes:

[0039] where X in is the input of this multi-scale dense block; Conv represents the convolution operation, and the subscripts 3×3, 5×5, 7×7, and 1×1 represent the convolution kernel sizes, and the superscript 1 represents the convolution stride; concat represents concatenation in the channel dimension; MaxPool represents max pooling, and the subscript 2×2 represents the pooling window size, and the superscript 2 represents the pooling stride; X 11 、X 12 、X13 , X 14 , X 21 , X 22 , X 23 and X 24 are feature maps generated during the operation process, and X out is the output of this multi-scale dense block. The output of the first multi-scale dense block is the input of the second multi-scale dense block.

[0040] The attention enhancement module is used to enhance the features of the second feature map by using the attention mechanism to obtain the third feature map. Specifically, the channel attention mechanism is used for feature enhancement, and the operation process of the attention enhancement module includes:

[0041] where x is the input of the attention enhancement module; GAP represents global average pooling; FC represents the fully connected layer; ReLu and Sigmoid are two activation functions; T is the channel weight; and y is the output of the attention enhancement module.

[0042] The shaping module is used to convert the third feature map into a feature vector. Specifically, the third feature map is flattened into a one-dimensional feature vector.

[0043] The classification module is used to map the feature vector to a probability vector of a preset scale by using a fully connected network. Specifically, three fully connected layers are used for classification, and the number of neurons is 128, 32, and 10 in sequence. The third fully connected layer uses the Softmax function to output the probabilities of each category, including 10 categories, namely normal, mild inner race fault, moderate inner race fault, mild outer race fault, moderate outer race fault, mild rolling element fault, moderate rolling element fault, mild compound fault, and moderate compound fault.

[0044] This embodiment proposes a three-branch network (multi-scale dense block) that performs multi-scale decomposition and feature extraction and fusion on the input data, thereby capturing more comprehensive and strongly correlated fault features, effectively enhancing the feature extraction and classification capabilities of the model. And the dense connection can also enhance the information flow and gradient propagation, and can alleviate the gradient disappearance, improve the training efficiency and convergence speed.

[0045] In one implementation, by introducing the channel attention mechanism, the importance of each channel is dynamically adjusted, redundant information is suppressed, and key fault features are highlighted, thereby improving the model recognition effect.

[0046] In one implementation, clarifying the fault level can help the operation and maintenance team reasonably arrange the maintenance priority, saving maintenance costs and time.

[0047] In one embodiment, the original data set is obtained by collecting one-dimensional vibration signals of the compressor under different loads, different rotational speeds, and different states, and overlapping sampling. The original data set is divided into a training set and a validation set. By performing random sampling with replacement on the original training set, training subsets for each state recognition model are obtained, allowing samples to be repeated. Multiple state recognition models are trained in parallel using different training subsets. The cross-entropy loss is used in the training process.

[0048] In one embodiment, the integration weights are determined by the training effects of multiple state recognition models. Specifically, after the training of multiple state recognition models is completed, the particle swarm optimization algorithm is used to optimize and determine the integration weights, including: Step 1: Use multiple state recognition models to predict each sample in the validation set to obtain the probability vector corresponding to each sample. For example, there are four state recognition models, and each model outputs a 10-dimensional probability vector. For each state category, there are 4 predicted probabilities, which are given by 4 state recognition models respectively.

[0049] Step 2: Initialize the population size, particle positions, and velocities. Specifically, the population size can be set to 30. Random numbers are generated within (0, 1) as particle positions. The particle positions are four-dimensional vectors, and the vector elements represent the fusion weights of the state recognition models. The velocities are initialized to 0.

[0050] Step 3: Use the Softmax function to convert the particle positions into effective weight distributions. The elements are normalized through the Softmax function so that the sum of the elements is 1.

[0051] Step 4: According to the effective weight distribution, integrate the probability vectors of each sample into probability scores to obtain the final prediction result, and calculate its accuracy as the fitness.

[0052] Step 5: Update the individual optimal position and the global optimal position according to the fitness value.

[0053] Step 6: Determine whether the preset stop condition is reached; if so, end the optimization process; otherwise, go to Step 7. Specifically, the stop condition includes reaching the maximum number of iterations. After ending the optimization process, use the Softmax function to normalize the global optimal position to obtain the optimal integration weights.

[0054] Step 7: Update the particle positions according to the individual optimal position and the global optimal position, and return to Step 3. Specifically, update the velocities and positions according to the rules of the particle swarm optimization algorithm, which will not be elaborated here.

[0055] In one implementation, by performing weighted fusion on the prediction results of multiple state recognition models and optimizing the weight allocation in combination with the performance on the validation set, the advantages of each model can be utilized more effectively, and finally a recognition accuracy better than that of a single model can be obtained.

[0056] In one implementation, the particle position is transformed into a probability distribution through the Softmax function, ensuring that all weights are non - negative and their sum is 1, avoiding meaningless negative values of weights or the situation where the sum exceeds 1, and improving the stability and interpretability of the fusion.

[0057] The embodiment of the present invention provides a safety monitoring system for the treatment of trichloroisocyanuric acid wastewater. Refer to Figure 4 , Figure 4 which is an architecture diagram of a safety monitoring system for the treatment of trichloroisocyanuric acid wastewater provided by the embodiment of the present invention. The system includes: A data acquisition module, configured to acquire the vibration signal of the compressor.

[0058] A pre - processing module, configured to pre - process the vibration signal to obtain a target signal.

[0059] A two - dimensional characterization module, configured to convert the target signal into a two - dimensional time - frequency image.

[0060] A base model decision module, configured to use the two - dimensional time - frequency image as the input of multiple pre - trained state recognition models to obtain the prediction probabilities of multiple state categories.

[0061] An integration module, configured to perform weighted summation on the prediction probabilities of the target state category according to the preset integration weights to obtain the probability score of the target state category.

[0062] A result output module, configured to output the state category corresponding to the maximum probability score as the detection result.

[0063] Wherein, the target state category is any one of the multiple state categories.

[0064] Based on the safety monitoring system for the treatment of trichloroisocyanuric acid wastewater provided by the embodiment of the present invention, by collecting the vibration signal of the compressor and converting it into a two - dimensional time - frequency image, using multiple pre - trained models for state recognition, and improving the prediction accuracy through weighted integration, early subtle faults can be effectively identified. Compared with the traditional threshold method, this method advances the fault recognition time point from the stage where the damage is obvious to the stage where the damage is not obvious, which helps to give early warnings and reduce the risk of sudden equipment failures.

[0065] In one embodiment, the pre - processing module includes: A segment extraction module, configured to extract a first vibration signal with a preset length from the vibration signal.

[0066] A modal decomposition module, which is used to perform variational modal decomposition on the first vibration signal to obtain a plurality of intrinsic mode components.

[0067] A correlation calculation module, which is used to calculate the Pearson correlation coefficient between each intrinsic mode component and the first vibration signal.

[0068] A signal reconstruction module, which is used to select the intrinsic mode components with the absolute value of the correlation coefficient greater than a preset threshold for signal reconstruction to obtain a target signal.

[0069] In one embodiment, the system further includes an integrated optimization module, which is used to optimize and determine the integrated weights by using a particle swarm optimization algorithm after multiple state recognition models are trained. The operation process of the integrated optimization module is the same as the process (steps one to seven) of optimizing and determining the integrated weights by using the particle swarm optimization algorithm described above, and will not be repeated here.

[0070] The above are only the preferred embodiments of the present invention, and do not impose any formal limitations on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to obtain equivalent embodiments with equivalent changes. However, as long as the technical content of the present invention is not departed from, any brief modifications, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A safety monitoring method for the treatment of trichloroisocyanuric acid wastewater, characterized in that, Treating wastewater by mechanical vapor recompression process, the method includes: Obtaining the vibration signal of the compressor and preprocessing the vibration signal to obtain a target signal; Converting the target signal into a two-dimensional time-frequency image; Taking the two-dimensional time-frequency image as the input of multiple pre-trained state recognition models to obtain the prediction probabilities of multiple state categories; According to the preset integration weights, weighted summing the prediction probabilities of the target state category to obtain the probability score of the target state category; the target state category is any one of the multiple state categories; Outputting the state category corresponding to the maximum probability score as the detection result.

2. The safety monitoring method for the treatment of trichloroisocyanuric acid wastewater according to claim 1, wherein, Preprocessing the vibration signal to obtain a target signal includes: Extracting a first vibration signal with a preset length from the vibration signal; Performing variational mode decomposition on the first vibration signal to obtain multiple intrinsic mode components; Calculating the Pearson correlation coefficient between each intrinsic mode component and the first vibration signal; Taking the intrinsic mode components with the absolute value of the correlation coefficient greater than the preset threshold for signal reconstruction to obtain the target signal.

3. A safety monitoring method for the treatment of trichloroisocyanuric acid wastewater according to claim 1, characterized in that, The structures of multiple state recognition models are the same; Any one state recognition model includes: An initial convolution module for performing shallow feature extraction on the two-dimensional time-frequency image using one convolutional layer and performing spatial compression to obtain a first feature map; A feature extraction module for performing multi-scale feature extraction on the first feature map using convolutional layers with different kernel sizes to obtain a second feature map; An attention enhancement module for enhancing the features of the second feature map using an attention mechanism to obtain a third feature map; A shaping module for converting the third feature map into a feature vector; A classification module for mapping the feature vector to a probability vector of a preset scale using a fully connected network.

4. A safety monitoring method for the treatment of trichloroisocyanuric acid wastewater according to claim 3, characterized in that, The feature extraction module includes two cascaded multi-scale dense blocks; The two multi-scale dense blocks have the same structure. The operation process of any one multi-scale dense block includes: Among them, X in is the input of this multi-scale dense block; Conv represents the convolution operation. The subscripts 3×3, 5×5, 7×7, and 1×1 represent the convolution kernel size, and the superscript 1 represents the convolution stride; concat represents concatenation in the channel dimension; MaxPool represents max pooling. The subscript 2×2 represents the pooling window size, and the superscript 2 represents the pooling stride; X 11 、X 12 、X 13 、X 14 、X 21 、X 22 、X 23 and X 24 are the feature maps generated during the operation process, and X out is the output of this multi-scale dense block.

5. The safety monitoring method for the treatment of trichloroisocyanuric acid wastewater according to claim 1, characterized in that, The integration weights are determined by the training effects of multiple state recognition models. Specifically, after multiple state recognition models are trained, a particle swarm optimization algorithm is used to optimize and determine the integration weights, including: Step 1, using multiple state recognition models to predict each sample in the validation set to obtain the probability vector corresponding to each sample; Step 2, initializing the population size, particle positions and velocities; Step 3, using the softmax function to convert the particle positions into an effective weight distribution; Step 4, according to the effective weight distribution, integrating the probability vectors of each sample into a probability score to obtain the final prediction result and calculating its accuracy as the fitness; Step 5, updating the individual optimal position and the global optimal position according to the fitness value; Step 6, determining whether the preset stop condition is reached; if so, ending the optimization process; otherwise, entering Step 7; Step 7, updating the particle positions according to the individual optimal position and the global optimal position, and returning to Step 3.

6. A safety monitoring system for the treatment of trichloroisocyanuric acid wastewater, which is used to implement a safety monitoring method for the treatment of trichloroisocyanuric acid wastewater as described in claim 1, and is characterized in that, The system includes: A data acquisition module for acquiring the vibration signal of the compressor; A preprocessing module for preprocessing the vibration signal to obtain a target signal; A two-dimensional characterization module for converting the target signal into a two-dimensional time-frequency image; The base model decision module is used to take the two-dimensional time-frequency image as the input of multiple pre-trained state recognition models, and obtain the prediction probabilities of multiple state categories; The integration module is used to sum the prediction probabilities of the target state category weighted according to the preset integration weights to obtain the probability score of the target state category; the target state category is any one of the multiple state categories; The result output module is used to output the state category corresponding to the maximum probability score as the detection result.

7. The safety monitoring system for trichloroisocyanuric acid wastewater treatment according to claim 6, characterized in that, The preprocessing module includes: The segment extraction module is used to extract the first vibration signal with a preset length from the vibration signal; The modal decomposition module is used to perform variational modal decomposition on the first vibration signal to obtain multiple intrinsic mode components; The correlation calculation module is used to calculate the Pearson correlation coefficient between each intrinsic mode component and the first vibration signal; The signal reconstruction module is used to take the intrinsic mode components with the absolute value of the correlation coefficient greater than the preset threshold for signal reconstruction to obtain the target signal.

8. The safety monitoring system for trichloroisocyanuric acid wastewater treatment according to claim 6, wherein, The structures of multiple state recognition models are the same; Any one of the state recognition models includes: The initial convolution module is used to perform shallow feature extraction on the two-dimensional time-frequency image by using one convolutional layer and perform spatial compression to obtain the first feature map; The feature extraction module is used to perform multi-scale feature extraction on the first feature map by using convolutional layers with different kernel sizes to obtain the second feature map; The attention enhancement module is used to perform feature enhancement on the second feature map by using the attention mechanism to obtain the third feature map; The shaping module is used to convert the third feature map into a feature vector; The classification module is used to map the feature vector to a probability vector with a preset scale by using a fully connected network.

9. The safety monitoring system for trichloroisocyanuric acid wastewater treatment according to claim 8, wherein The feature extraction module includes two cascaded multi-scale dense blocks; The two multi-scale dense blocks have the same structure. The operation process of any one multi-scale dense block includes: where X in is the input of this multi-scale dense block; Conv represents the convolution operation, and the subscripts 3×3, 5×5, 7×7, and 1×1 represent the convolution kernel size, and the superscript 1 represents the convolution stride; concat represents concatenation in the channel dimension; MaxPool represents max pooling, and the subscript 2×2 represents the pooling window size, and the superscript 2 represents the pooling stride; X 11 、X 12 、X 13 、X 14 、X 21 、X 22 、X 23 and X 24 are the feature maps generated during the operation process, and X out is the output of this multi-scale dense block.

10. The safety monitoring system for trichloroisocyanuric acid wastewater treatment according to claim 6, characterized in that, The system further includes an integration optimization module, which is used to optimize and determine the integration weights by using the particle swarm optimization algorithm after multiple state recognition models are trained; the operation process of the integration optimization module includes: Step 1, use multiple state recognition models to predict each sample in the validation set to obtain the probability vector corresponding to each sample; Step 2, initialize the population size, particle position and velocity; Step 3, use the softmax function to convert the particle position into an effective weight distribution; Step 4, according to the effective weight distribution, integrate the probability vectors of each sample into a probability score to obtain the final prediction result, and calculate its accuracy as the fitness; Step 5, update the individual optimal position and the global optimal position according to the fitness value; Step 6, determine whether the preset stop condition is reached; if so, end the optimization process; otherwise, go to Step 7; Step 7, update the particle position according to the individual optimal position and the global optimal position, and return to Step 3.

Images