A method for intelligent monitoring and early warning of odor in household garbage collection process

Abstract

The present application relates to a kind of household garbage collection process odor intelligent monitoring early warning method, realize the prediction of odor concentration, and determine early warning level, realize odor intelligent monitoring early warning, comprising the following steps: first, data is collected and data is preprocessed;Determine the input variable and output variable of model;Then, using modular neural network establishes odor concentration prediction model;Finally, reference is made to "Standard for Odor Pollutants Emission (GB 14554-93)" and "Indoor Air Quality Standard (GB / T 18883-2002)", set hydrogen sulfide H2S, ammonia NH3 concentration early warning threshold, draft odor early warning level, realize odor monitoring early warning.The present application effectively realizes the accurate prediction of odor concentration and monitoring early warning, and has important theoretical significance and application value.

Description

TECHNICAL FIELD

[0001] The present application relates to a method for intelligent monitoring and early warning of household garbage collection process; a hydrogen sulfide H2S and ammonia NH3 prediction model based on modular neural network is established to realize intelligent prediction of H2S and NH3 concentration trend; by setting H2S and NH3 concentration early warning threshold, odor early warning level is drafted to realize odor monitoring and early warning. It belongs to both the field of household garbage treatment and the field of intelligent modeling. BACKGROUND

[0002] With the rapid development of China's economy and the continuous acceleration of urbanization process, the amount of urban household garbage is increasing day by day. In the process of household garbage collection, the types of garbage are various and the composition is complex, which is easy to cause pollution, especially perishable garbage. Perishable garbage accounts for a large proportion in household garbage, has high degradability and is easy to rot, which is the source of malodorous gas in the process of household garbage collection, and seriously endangers the physical and mental health of residents and the living environment. H2S and NH3 are the main components of malodorous gas, and their concentration trend is predicted and the early warning level is determined, so that appropriate measures can be taken to control odor emission. Therefore, it has important theoretical significance and application value to realize the monitoring and early warning of odor concentration. SUMMARY

[0003] The purpose of the present application is to provide a method for intelligent monitoring and early warning of household garbage collection process, which uses modular neural network to establish H2S and NH3 concentration prediction model to realize prediction of odor concentration at future time. By setting H2S and NH3 concentration early warning threshold, odor early warning level is drafted to realize odor monitoring and early warning.

[0004] The present application adopts the following technical solutions and implementation steps:

[0005] 1. Collecting field data;

[0006] Real-time collection and storage of hydrogen sulfide H2S and ammonia NH3 concentration data by equipment installed in household garbage classification post, sampling interval time is 30 seconds;

[0007] Step 2, determine the input and output variables of the prediction model;

[0008] The input variables of the prediction model are represented as x=(x1,x2,x3,x4) T , which respectively represent H2S and NH3 concentration at t-3, t-2, t-1 and t time, and the output variable y of the model is H2S and NH3 concentration at t+2 time;

[0009] Step 3, design modular neural network and establish prediction model;

[0010] The method for establishing the prediction model is as follows: first, modular partition is realized based on the attention mechanism; second, a subnetwork is constructed based on a self-organizing fuzzy neural network; and finally, the outputs of each subnetwork are integrated based on the attention mechanism to obtain a final prediction result, as follows:

[0011] (1) The modular partition method based on the attention mechanism is as follows:

[0012] The modular index MQ is composed of the density within the module and the sparsity between the modules, wherein the density within the module is calculated as follows:

[0013]

[0014] Wherein, J C is the density within the module, P is the number of modules in the current network, N l is the number of samples allocated to the lth module, x i is the input vector of the ith training sample, h l and r l respectively represent the core node and the action range of the lth module.

[0015] The sparsity between the modules is calculated based on the Euclidean distance:

[0016]

[0017] Wherein, J S is the sparsity between the modules, d(h l , h s ) represents the distance between the core node of the lth module and the core node of the sth module, according to the density within the module J C and the sparsity between the modules J S , the modular index is proposed as follows:

[0018]

[0019] The greater the value of MQ, the higher the modular degree of the network, and the modular partition method based on the attention mechanism is as follows: first, the core nodes of each initial sub-module are determined according to the point density; second, samples are allocated to each sub-module based on the attention mechanism; then, the new core nodes of the existing modules are determined by seeking the maximum "modularity" of the network; finally, the optimal number of modules in the network is determined by merging the nearest core nodes and seeking the maximum "modularity" of the network.

[0020] ① Determining the center of each initial sub-module

[0021] The data set cannot explicitly describe the sample point distribution, and the point density idea is introduced, that is, the density of sample point distribution. The more points around the sample point, the greater the sample distribution density at the sample point, and the more likely the sample becomes the core node of the modular partition. Therefore, a point density function is proposed as a method for selecting the initial sub-module core node.

[0022] For each training sample input vector x, the expression of the point density function is defined as follows:

[0023]

[0024] d ij ||x i -x j || (5)

[0025] Where d ij represents the distance between two training sample input vectors x i and x j ; the more points around x i , the greater the value of Z i ;

[0026] Calculate the point density of each sample, and take the input vectors of the first n training samples with the maximum point density as the core nodes of the initial sub-module, which is specifically set as follows:

[0027]

[0028] Where h i and r i represent the core node and the range of action of the i-th initial sub-module, respectively; x Zi is the i-th training sample input vector with the maximum point density; d max is the maximum distance between the training sample input vectors x i and x j .

[0029] ②Based on the attention mechanism to allocate samples

[0030] At time T, when the T-th training sample x T enters the network, calculate the attention value of each sub-module to x T :

[0031]

[0032] Where Attention d is the attention value of the d-th sub-module to x T , Q d is the core node of the d-th sub-module, and Q diLet x be the i-th component of the d-th submodule core node, and K and V be the input sample x. T As shown in equation (10), K i V i For sample x T The i-th component; S i The correlation value is obtained as shown in equation (11);

[0033] K = V = x T =[x1,…,x i ,…,x k (10)

[0034] Submodule d on sample x T The specific steps for calculating the attention value are as follows:

[0035] a) According to Q d The correlation between the two is calculated using the formula:

[0036] S i =|Q di -K i | (11)

[0037] b) For S i The value is normalized using softmax to obtain a. i value;

[0038] c) Based on a obtained in step b), i Value calculation 1-a i value;

[0039] d) Based on the obtained 1-a i The attention value is obtained by weighting and summing the values ​​of V. d Complete submodule d for x T Attention value calculation;

[0040] If the d-th submodule is related to x T If the attention value is greater than the threshold λ, i.e., Attention ≥ λ, then x will be... T Assigned to this module, λ is set to 0.5;

[0041] ③ Optimization module

[0042] To achieve optimal modularity in the network, the modularity index MQ of the entire network is calculated according to formula (3), with the newly added samples and the original module core nodes serving as sub-module core nodes respectively. t and MQ lmin :

[0043] If MQ lmin <MQ t, the network modularity degree of the current new sample as the core node is higher, and the sample is replaced by the original module core node to become a new module core node, and the initial parameter setting is as follows:

[0044] h t = x t (12)

[0045]

[0046] where h t is the new core node of the module, x t is the input vector of the new sample, r t is the action range of the module, and N lmin is the number of samples allocated to the module;

[0047] If MQ lmin > MQ t , the current module core node remains unchanged, and only the action range of the module needs to be adjusted:

[0048]

[0049] where r lmin is the action range of the module, and N lmin is the number of samples allocated to the module;

[0050] After all the training samples are compared, the samples are allocated to different sub-modules;

[0051] ④ Merge module

[0052] If at time T, there are two or more modules, calculate the Euclidean distance between the core nodes of each module, and when the distance between core nodes c i and c j is the smallest, the two modules are combined into a new module, and the initial parameter setting is as follows:

[0053]

[0054] where c new is the core node of the merged module, Z ci and Z cj are the point densities of core nodes c i and c j , respectively;

[0055] According to formula (3), the modularity index value MQ of the entire network before and after merging the module is calculated, and the number of modules when MQ value is maximum is found. At this time, the modularity degree of the network is the largest, and the modularity partition is formed;

[0056] (2) A self-organizing fuzzy neural network is used to construct the sub-network corresponding to each module. The design of the sub-network includes three parts: network structure growth, network structure pruning and network parameter adjustment.

[0057] ① Network structure growth

[0058] When the first training sample enters the network, this sample is used as the center, width, and connection weights to the output layer of the first RBF neuron:

[0059] c1 = x1 (16)

[0060] σ1=1 (17)

[0061] w1 = y d1 (18)

[0062] Where c1 is the center of the first neuron, x1 is the input vector of the first training sample, σ1 is the width of the first neuron, w1 is the connection weight from the regularization layer neuron to the output layer neuron, and y d1 This is the output value of the first training sample;

[0063] At time t s At time t, the training error vector e(t) s It is obtained from the following formula:

[0064]

[0065] Among them, y f It is the expected output value of the f-th training sample. It is the f-th sample at time t s The network output value;

[0066] Find the sample where the expected output value and the network output value differ the most.

[0067]

[0068] Add a new rule for the first We learn from each sample to compensate for the current maximum error; the center vector and output weights of the newly added rule are set as follows:

[0069] c t =x l (twenty two)

[0070] w t =y dl (twenty three)

[0071] Among them, c t x is the center of the newly added neuron. lis the input vector of the lth training sample, w t is the connection weight from the regularized layer neuron to the output layer neuron, y dl is the output value of the lth training sample;

[0072] After adding a new rule, the Euclidean distance between the new rule and the existing rules is calculated and the minimum distance d min :

[0073] d min = min{dist(c t ,c j≠t )} (24)

[0074] Based on the minimum distance, the width σ t of the new neuron is set as:

[0075] σ t = 0.7d min (25)

[0076] Each time a new rule is added, the network parameters are adjusted by a second-order learning algorithm; when the preset maximum structure J max or the desired training accuracy E0 is reached, the network structure growth process ends; in the experiment, J max = 10, E0 = 0.0001, and the sum of squared errors SSE is used to measure the training accuracy of the network, which is calculated as follows:

[0077]

[0078] where y i and are the expected output value and the network output value of the ith training sample, respectively, and N s is the number of samples in the sub-module s;

[0079] ② Network structure pruning

[0080] To avoid redundancy in the network structure, a metric value based on the activation strength of the RBF layer neurons is proposed:

[0081]

[0082] where Φ(j) is the output of the jth RBF neuron, and J min s is the neuron with the minimum activation strength in the RBF layer;

[0083] Therefore, the rule with the minimum activation strength is deleted, and the root mean square error value SSE_1 after deleting the rule is compared with the root mean square error value SSE_0 before deleting the rule by adjusting the parameters through the second-order learning algorithm; if SSE_1 ≤ SSE_0, the selected rule can be pruned without sacrificing the learning ability of the network, then the above process is repeated, and conversely, the selected rule cannot be deleted; at this time, the network structure pruning process is completed, and the sub-network construction is completed; each time the neurons are deleted, the parameters are adjusted using the second-order learning algorithm;

[0084] ③Network parameter adjustment

[0085] The improved second-order learning algorithm is as follows:

[0086] Δ t+1 =Δ t -(Q t +μ t I) -1 g t (28)

[0087] Where Δ t+1 is the adjusted parameter, Δ is the parameter before adjustment, including the center, width and connection weight; Q t is the quasi-Hessian matrix; g t is the gradient vector; I is the unit matrix; μ t is the learning rate parameter, which is 0.01.

[0088] In order to reduce the space complexity and time complexity in the calculation process of the Jacobian matrix, the calculation of the quasi-Hessian matrix Q t and the gradient vector g t is converted into the summation of the sub-quasi-Hessian matrix and the summation of the sub-gradient vector:

[0089]

[0090] q z is the sub-quasi-Hessian matrix, η z is the sub-gradient vector, and both can be calculated by the following formula:

[0091]

[0092] Where e z is the difference between the expected output value y z of the zth sample and the network prediction output value , j z is the Jacobian vector, and is calculated as follows:

[0093]

[0094] According to the chain rule, each component in the Jacobian vector in equation (34) is calculated as follows:

[0095]

[0096] where c j , σ j , w j are the center, width and connection weight to the output layer of the jth neuron, respectively, and x z is the input vector of the zth training sample;

[0097] (3) The integration strategy based on the attention mechanism is as follows:

[0098] The test data is input to the model, and the output of the model is the predicted value of the H2S and NH3 concentration at time t+2:

[0099] At time T, when the Tth test sample enters the model, the attention value Attention of each submodule to x T is calculated, and when Attention≥λ, the corresponding submodule is activated to calculate the input sample;

[0100] After each submodule calculates the sample, the output integration layer integrates the outputs of each activated submodule:

[0101]

[0102] where y r is the output value of the rth subnetwork, and ε r is the weight coefficient of the rth subnetwork to the output layer, which is determined according to the attention degree of the subnetwork to the input sample, and the calculation formula is as follows:

[0103]

[0104] where Attention r is the attention value of the rth activated subnetwork to the input sample x T , and R is the total number of activated subnetworks.

[0105] Step 4, set the H2S and NH3 concentration warning thresholds, and draft the odor warning levels to realize odor warning;

[0106] According to the "Standard for Emission of Odorous Pollutants (GB 14554-93)" and "Indoor Air Quality Standard (GB / T 18883-2002)", the H2S and NH3 concentration warning thresholds are set, and the proposed odor warning levels are shown in Table 1. When the H2S and NH3 concentrations reach the corresponding thresholds, the corresponding level of warning is performed, and the corresponding level of control measures is taken to reduce the emission of H2S and NH3.

[0107] Table 1: H2S and NH3 concentration warning levels

[0108]

[0109]

[0110] The root mean square error (RMSE) and the mean absolute percentage error (MAPE) are used to quantitatively evaluate the test accuracy, and the RMSE and MAPE calculation formulas are as follows:

[0111]

[0112] In the formula, y i and are the expected output and network output of the i-th sample respectively, and N T is the number of test samples.

[0113] The present application is based on the good nonlinear mapping ability and generalization ability of the modular neural network, and a stable and effective odor concentration prediction model is established, which can realize accurate prediction of odor concentration, and through determination of the warning level of odor concentration, corresponding measures are taken to control the emission of odor concentration, which has important significance for odor emission control in the process of collecting household garbage. BRIEF DESCRIPTION OF DRAWINGS

[0114] Figure 1 is a flowchart of the present application;

[0115] Figure 2 is a basic structure diagram of the modular neural network;

[0116] Figure 3 is a self-organizing fuzzy neural network structure diagram;

[0117] Figure 4 is an attention mechanism calculation flowchart;

[0118] Figure 5 is an integration flowchart of the subnetwork based on the attention mechanism;

[0119] Figure 6 is a H2S concentration test result diagram;

[0120] Figure 7 is a H2S concentration test error diagram;

[0121] Figure 8 is a NH3 concentration test result diagram;

[0122] Figure 9 is a NH3 concentration test error diagram. DETAILED DESCRIPTION

[0123] The application utilizes a training data set to establish a modular neural network model for odor concentration prediction; and utilizes a test data set to verify the accuracy of the future time odor concentration prediction value output by the modular neural network model.

[0124] As an embodiment, the effectiveness of the method is verified by using data from a waste sorting station in Beijing, 2800 groups of data on June 15, 2021 are selected for experiment, of which the first 2000 groups of data are used as training data, and the remaining 800 groups of data are used as test data.

[0125] (1) Based on 2000 groups of training data, the odor concentration at t-3, t-2, t-1 and t time is selected as the model input, and the odor concentration at t+2 time is selected as the model output, to establish an odor concentration prediction model based on modular neural network;

[0126] (2) Based on 800 groups of test data, odor concentration prediction is carried out by modular neural network, and the test results of H2S and NH3 are shown in Figure 6 、 Figure 8 , X axis: test sample number, unit: mg / sample, Y axis: odor concentration, unit: mg / m 3 ; the test error is shown in Figure 7 、 Figure 9 , X axis: test sample number, unit: mg / sample, Y axis: odor concentration test error, unit: mg / m 3 ;

[0127] (3) The root mean square error RMSE and the average percentage error MAPE are used to quantitatively evaluate the test accuracy, the calculation results of H2S concentration are RMSE=0.0037 and MAPE=9.8259%, and the calculation results of NH3 concentration are RMSE=0.0139 and MAPE=0.8006%;

[0128] (4) Based on the prediction results of odor concentration, and combined with the odor warning level table, the odor warning level is judged, and corresponding measures are taken to control odor emission.

[0129] Obviously, those skilled in the art can make various modifications and changes to the application without departing from the spirit and scope of the application. It should be noted that the above description is only a specific embodiment of the application and does not limit the application, and any modification and optimization made within the spirit and principles of the application shall be covered by the scope of the claims of the application.

Claims

1. A method for intelligent monitoring and early warning of odor during the collection of domestic waste, characterized in that, Includes the following steps: Step 1: Collect on-site data; The equipment installed at the household waste sorting station collects and stores data on the concentrations of hydrogen sulfide (H2S) and ammonia (NH3) in real time, with a sampling interval of 30 seconds. Step 2: Determine the input and output variables of the prediction model; The input variables for the prediction model are represented as x = (x1, x2, x3, x4). T , representing the H2S and NH3 concentrations at times t-3, t-2, t-1, and t, respectively, and the model's output variable y is the H2S and NH3 concentration at time t+2; Step 3: Design a modular neural network, establish a prediction model, and realize the prediction of H2S and NH3 concentrations; Step 4: Set H2S and NH3 concentration warning thresholds, determine odor warning levels, and implement odor warning; In step 3, the method for designing a modular neural network and establishing a prediction model is as follows: First, modular partitioning is implemented based on an attention mechanism; second, sub-networks are constructed based on a self-organizing fuzzy neural network; finally, the outputs of each sub-network are integrated based on an attention mechanism to obtain the final prediction result, as detailed below: (1) The modular partitioning method based on the attention mechanism is as follows: The modularity metric MQ is composed of the density within modules and the sparsity between modules. The density within modules is calculated as follows: Among them, J C The density within a module, P is the number of modules in the current network, and N is the density within the module. l x is the number of samples assigned to the l-th module. i Let h be the input vector of the i-th training sample. l and r l These represent the core node and scope of the l-th module, respectively. The sparsity between modules is calculated based on Euclidean distance: Among them, J S To represent the sparsity between modules, d(h) l ,h s J represents the distance between the core nodes of the l-th and s-th modules, depending on the density within the modules. C Sparsity J between modules S The following modular indicator measurement method is proposed: The larger the MQ value, the higher the modularity of the network. The modular partitioning method based on the attention mechanism is as follows: First, determine the core nodes of each initial sub-module based on the idea of ​​point density; second, allocate samples to each sub-module based on the attention mechanism; then, determine the new core nodes of the existing modules by seeking the maximum "modularity" of the network; finally, determine the optimal number of modules in the network by merging the nearest core nodes and seeking the maximum "modularity" of the network. ① Determine the center of each initial submodule Since the dataset cannot explicitly describe the distribution of sample points, the concept of point density is introduced, which is the density of the sample point distribution. If there are more points around a sample point, the sample distribution density at that sample point is greater, and the sample is more likely to become the core node of the modular partition. Therefore, a point density function is proposed as a method for selecting the core node of the initial submodule. For each training sample's input vector x, its point density function is defined as follows: d ij =||x i -x j || (5) Where, d ij Represents two training sample input vectors x i and x j The distance between; if x i The more points surrounding it, the more Z... i The larger the value; Calculate the point density of each sample, and use the input vectors of the n training samples with the highest point density as the core nodes of the initial submodule, as follows: Among them, h i and r i x represents the core node and scope of the i-th initial submodule, respectively; Zi d is the input vector of the training sample with the highest density at the i-th point; max Input vector x for training samples i and x j The maximum distance between; ② Sample allocation based on attention mechanism At time T, when the T-th training sample x T When entering the network, calculate the relationship between each submodule and x. T Attention value: Among them, Attention d For the d-th submodule, x T Attention value, Q d Q is the core node of the d-th submodule. di Let x be the i-th component of the d-th submodule core node, and K and V be the input sample x. T As shown in equation (10), K i V i For sample x T The i-th component; S i The correlation value is obtained as shown in equation (11); K=V=x T =[x1,…,x i ,…,x k ] (10) Submodule d on sample x T The specific steps for calculating the attention value are as follows: a) According to Q d The correlation between the two is calculated using the formula: S i =|Q di -K i | (11) b) For S i The value is normalized using softmax to obtain a. i value; c) Based on a obtained in step b), i Value calculation 1-a i value; d) Based on the obtained 1-a i The attention value is obtained by weighting and summing the values ​​of V. d Complete submodule d for x T Attention value calculation; If the d-th submodule is related to x T If the attention value is greater than the threshold λ, i.e., Attention ≥ λ, then x will be... T Assigned to this module, λ is set to 0.5; ③ Optimization module To achieve optimal modularity in the network, the modularity index MQ of the entire network is calculated according to formula (3), with the newly added samples and the original module core nodes serving as sub-module core nodes respectively. t and MQ lmin : If MQ lmin <MQ t Using the newly added sample as the core node results in a higher degree of network modularity. This sample replaces the original module core node to become the new module core node, with the initial parameter settings as follows: h t =x t (12) Among them, h t x is the new core node of the module. t For the input vector of the newly added sample, r t N represents the scope of the module. lmin This represents the number of samples assigned to this module. If MQ lmin >MQ t The core nodes of the current module remain unchanged; only the scope of the module needs to be adjusted. Where, r lmin N represents the scope of the module. lmin This represents the number of samples assigned to this module. After all training samples have been compared, the samples are assigned to different sub-modules; ④ Merging Module If at time T there are two or more modules, calculate the Euclidean distance between the core nodes of each module. When the core node c i to c j If the distance between the two modules is the smallest, then the two modules will be combined into a new module, with the initial parameters set as follows: Among them, c new Z is the core node of the merged module. ci and Z cj These are the core nodes c. i and c j Point density; According to formula (3), calculate the modularity index value MQ of the entire network before and after merging modules, and find the number of modules when the MQ value is the largest. At this time, the modularity of the network is the largest, and modular partitioning is formed. (2) A self-organizing fuzzy neural network is used to construct the sub-network corresponding to each module. The design of the sub-network includes three parts: network structure growth, network structure pruning and network parameter adjustment. ① Network structure growth When the first training sample enters the network, this sample is used as the center, width, and connection weights to the output layer of the first RBF neuron: c1 = x1 (16) σ1=1 (17) w1=y d1 (18) Where c1 is the center of the first neuron, x1 is the input vector of the first training sample, σ1 is the width of the first neuron, w1 is the connection weight from the regularization layer neuron to the output layer neuron, and y d1 This is the output value of the first training sample; At time t s At time t, the training error vector e(t) s It is obtained from the following formula: Among them, y f It is the expected output value of the f-th training sample. It is the f-th sample at time t s The network output value; Find the sample where the expected output value and the network output value differ the most. Add a new rule for the first We learn from each sample to compensate for the current maximum error; the center vector and output weights of the newly added rule are set as follows: c t =x l (22) w t =y dl (23) Among them, c t x is the center of the newly added neuron. l Let w be the input vector of the l-th training sample. t y represents the connection weights from neurons in the regularization layer to neurons in the output layer. dl This is the output value of the l-th training sample; After each new rule is added, the Euclidean distance between the new rule and existing rules is calculated, and the minimum distance d is found. min : d min =min{dist(c t ,c j≠t )} (24) Based on this minimum distance, the width σ of the newly added neuron will be... t Set to: s t =0.7d min (25) Each time a new rule is added, the network parameters are adjusted using a second-order learning algorithm; when the preset maximum structure J is reached... max Or, if the desired training accuracy is E0, the network structure growth process ends; during the experiment, J... max =10, E0=0.0001, the sum of squared errors (SSE) is used to measure the training accuracy of the network, and is calculated as follows: Among them, y i and Let N be the expected output value and the network output value of the i-th training sample, respectively. s The number of samples in submodule s; ② Network structure pruning To avoid redundancy in the network structure, a metric based on the activation intensity of neurons in the RBF layer is proposed: Where Φ(j) is the output of the j-th RBF neuron, J min s The neuron with the lowest activation intensity in the RBF layer; Therefore, the rule with the lowest activation strength is deleted, and the parameters are adjusted using a second-order learning algorithm. The root mean square error (RMSE) value SSE_1 after deleting the rule is compared with the RMSE value SSE_0 before deleting the rule. If SSE_1 ≤ SSE_0, the selected rule can be pruned without sacrificing the network's learning ability. Then, the above process is repeated. Otherwise, the selected rule cannot be deleted. At this point, the network structure pruning process ends, and the subnetwork construction is completed. The parameters are adjusted using a second-order learning algorithm each time a neuron is deleted. ③ Network parameter adjustment The improved second-order learning algorithm is shown below: D t+1 =D t -(Q t +m t I) -1 g t (28) Where, Δ t+1 The adjusted parameter is Δ. t The parameters before adjustment include center, width, and connection weights; Q t For Hessian-like matrices; g t I is the gradient vector; I is the identity matrix; μ t The learning rate parameter is set to 0.

01. To reduce the space and time complexity of the Jacobian matrix calculation, the Hessian-like matrix Q is... t and gradient vector g t The calculation is transformed into summation of subclass Hessian matrices and summation of sub-gradient vectors: q z For subclass Hessian matrix, η z The sub-gradient vectors can all be calculated using the following formula: Among them, e z The expected output value y for the z-th sample z Compared with network prediction output value The difference, j z Let the Jacobian vector be denoted as follows: According to the chain rule, each component of the Jacobian vector in equation (34) is calculated as follows: Where c j , σ j w j x represents the center, width, and connection weights to the output layer of the j-th neuron, respectively. z It is the input vector of the z-th training sample; (3) The integration strategy based on the attention mechanism is as follows: The test data is used as input to the model, and the model output is the predicted concentrations of H2S and NH3 at time t+2. At time T, after the Tth test sample enters the model, the relationship between each submodule and x is calculated. T The attention value is set to λ. When Attention ≥ λ, the corresponding sub-module is activated to calculate the input sample. After each submodule has performed calculations on the samples, the output integration layer performs a weighted integration of the outputs of each activated submodule: Among them, y r Let ε be the output value of the r-th subnetwork. r Let be the weight coefficient from the r-th subnetwork to the output layer. It is determined based on the degree of attention the subnetwork pays to the input sample, and the calculation formula is as follows: Among them, Attention r For the r-th activated subnetwork to input sample x T The attention value; R is the total number of activated subnetworks.

2. The intelligent odor monitoring and early warning method for the domestic waste collection process as described in claim 1, characterized in that, In step 4, H2S and NH3 concentration warning thresholds are set, odor warning levels are determined, and odor warning is implemented. Referring to the "Odor Pollutant Emission Standard (GB 14554-93)" and the "Indoor Air Quality Standard (GB / T 18883-2002)," warning thresholds for H2S and NH3 concentrations are set, and the proposed odor warning levels are as follows: (1) H2S concentration prediction levels: Level 1 warning when m < 0.03; Level 2 warning when 0.03 ≤ m < 0.06; m ≥ 0.06, Level III warning; m is the H2S concentration, unit is mg / m³. 3 ; (2) NH3 concentration warning levels: n<0.2, Level 1 warning; 0.2≤n<1.0, Level 2 warning; 1.0≤n<1.5, Level 3 warning; n≥1.5, Level 4 warning; n is the NH3 concentration, in mg / m³. 3 ; When the concentrations of H2S and NH3 reach the corresponding thresholds, an early warning of the corresponding level will be issued, and corresponding control measures will be taken to reduce H2S and NH3 emissions.

Images