A production line energy consumption data management system and method based on deep learning

Abstract

The present invention discloses a production line energy consumption data management system and method based on deep learning, which relates to the technical field of energy consumption data management. The present invention collects the three-dimensional positioning coordinates of production equipment and calibrates them, extracts the expanded time window corresponding to historical fault events; analyzes the spatial distribution density index between production equipment when a production equipment failure occurs, and establishes a pre-control model for the production equipment based on the filtered and processed data; monitors the operation of production equipment in the current space, issues an alarm when abnormal operation of production equipment is detected, and selectively transmits the monitoring data in the current space; retrains the pre-control model based on the selectively transmitted data set, and dynamically adjusts the hidden layer dimension of the pre-control model to avoid data omission and improve the accuracy of subsequent analysis; and adaptively adjusts the alarm sensitivity to reduce redundant data transmission.

Description

Technical Field

[0001] The present invention relates to the technical field of energy consumption data management, and specifically to a production line energy consumption data management system and method based on deep learning. Background Art

[0002] Currently, my country's industry is still generally faced with the challenges of high investment, high consumption, and high emissions. Unregulated and disorderly energy use has resulted in suboptimal energy efficiency and high energy costs for enterprises, which runs counter to environmental, social, and governance goals and the "dual carbon" initiative. Therefore, promoting green and intelligent industrial development and reducing industrial energy consumption have become inevitable trends in my country's industrial development.

[0003] Existing data collection and transmission mechanisms suffer from redundancy and inefficiency. Traditional monitoring systems typically collect full sensor data at a fixed frequency, resulting in a large amount of irrelevant or low-value information occupying communication bandwidth and storage resources. In anomaly detection scenarios, redundant data not only increases the computational burden but can also obscure key features, hindering the efficiency of real-time decision-making.

[0004] Therefore, the present invention discloses a production line energy consumption data management system and method based on deep learning to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a production line energy consumption data management system and method based on deep learning to solve the problems raised in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a production line energy consumption data management method based on deep learning, the method comprising the following steps:

[0007] S1: Collect and calibrate the three-dimensional positioning coordinates of production equipment, extract the extended time window corresponding to historical fault events, and integrate the spatial distribution of the production equipment location and operating status before and after the fault;

[0008] S2: Analyze the spatial distribution density index between production equipment when a production equipment failure occurs, filter and process the historical data to be regulated, and establish a pre-regulation model for the production equipment based on the filtered data;

[0009] S3: Monitor the operation of production equipment in the current space, issue an alarm when abnormal operation of production equipment is detected, and selectively transmit the monitoring data in the current space;

[0010] S4: Retrain the pre-conditioning model based on the selectively transmitted dataset and dynamically adjust the hidden layer dimensions of the pre-conditioning model.

[0011] According to the above scheme, S1 includes the following contents:

[0012] S101: Obtain the real-time positioning data of each production equipment, generate the original coordinates with timestamps, apply Kalman filtering to the original coordinates to generate the standardized coordinate set D = {(x i ,y i , z i , t i )|i∈[1,I], i is a positive integer}; where (x i ,y i , z i ) represents the three-dimensional coordinates of the i-th production equipment, t i represents the timestamp corresponding to the three-dimensional coordinates of the i-th production equipment, and I represents the total number of production equipment in the space to be analyzed;

[0013] S102: Extract fault event data from the historical database and generate a fault event set E={e m |m∈[1, M], where m is a positive integer}; each fault event contains the faulty production equipment number, the set of affected production equipment, the set of control parameters, and the time interval of the fault event; M represents the total number of fault events;

[0014] S103: Set the fault event time interval [t m start ,t m end ], mapped to the time axis of the standardized coordinate set D, based on the preset event window △t, the fault event time interval is expanded to generate the event retrieval time interval, recorded as T m ={t|t∈[t m start -△t,t m end +△t]}, t represents the time; for the fault event e m In the event retrieval time interval T m The coordinates and distribution status of the corresponding production equipment are retrieved and recorded as the fault event set E'={(e m , U m )|m∈[1,M]}; where U m Indicates fault event e m The production equipment location distribution status, U m ={(x i ,y i , z i ,t,S （i，m） )|i∈[1,I], i is a positive integer}; where S （i，m） Indicates fault event e mThe operating status code of the production equipment i when it occurs; the operating status code includes the real-time load rate and working mode, and the working mode includes on and off.

[0015] In this application, Karl uses Mann filtering to standardize coordinates to eliminate sensor noise and improve the spatiotemporal consistency of device location data. By dynamically expanding the fault event time window to cover the fault incubation period and recovery period, data omissions are avoided and the accuracy of subsequent analysis is improved.

[0016] According to the above scheme, S2 includes the following contents:

[0017] S201: Analyze the distance between each production equipment when the fault occurs based on the fault event set with spatial labels; classify the production equipment whose distance to production equipment i is less than the dynamic influence radius of production equipment i into the neighborhood production equipment set of production equipment i, recorded as N i The dynamic influence radius is equal to the product of the real-time load rate and the influence coefficient, and the influence coefficient is a system preset constant; Analyze the neighborhood production equipment set N i The closeness index COM i :

[0018] ;

[0019] Among them, n represents the serial number of the production equipment in the neighborhood production equipment set, d （i，n） represents the distance between production equipment n and production equipment i in the set of neighboring production equipment of production equipment i, σ represents the spatial coefficient; exp() represents an exponential function with natural numbers as the base; the spatial coefficient is equal to the product of the spatial attenuation coefficient and the average spacing, the spatial attenuation coefficient is a system preset constant, and the average spacing is the average value of the spacing between production equipment in the monitoring space; analyze the compactness index of all production equipment in the monitoring space and perform normalization processing;

[0020] S202: Analyze the similarity between any two production equipment and construct a similarity matrix A; record the similarity between production equipment i and production equipment j in the monitoring space as A （i，j） ;

[0021] ;

[0022] Among them, d （i，j） It represents the distance between production equipment i and production equipment j in the monitoring space;

[0023] Based on the similarity matrix A, a degree matrix D is constructed. The degree matrix D is a diagonal matrix. （j，j） =∑ i A （i，j）Based on the similarity matrix A and the degree matrix D, a symmetric normalized Laplace matrix is ​​generated. The eigenvector of the Laplace matrix is ​​solved based on the spectral clustering algorithm. The optimal number of clusters is determined by the gap statistics method. The production equipment cluster division result {C1, ..., C Q}; where Q represents the total number of production equipment clusters;

[0024] S203: Construct a pre-regulation model, which is a two-layer LSTM network model including a cluster coordination layer and a device execution layer, wherein the cluster coordination layer network of the two-layer LSTM network model receives the cluster state matrix VT t ={VT q t |q∈[1,Q], q is a positive integer} as input; VT q t Represents the production equipment cluster C q Vibration temperature fusion characteristics; VT q t =avg[COM k , VIB k , TEM k ] t , k∈C q Among them, VIB k represents the vibration frequency of the kth production equipment, TEM k Represents the temperature of the k-th production equipment, avg[] represents the average function; the inter-cluster control weight matrix W is output through the time attention mechanism t ;

[0025] The bidirectional LSTM layer in the cluster coordination layer is composed of two stacked layers. The number of hidden units in each direction is 64. The activation function is the GRU activation function instead of the traditional LSTM to reduce the number of parameters. The input dimension is Q×3 and the output dimension is Q×128.

[0026] Attention weight a in the temporal attention layer t =Softmax(UV T / (dk) 1 / 2 ); where U=VT t ×W U , V=VT t ×W V ;W U and W V Represents the learning parameter, dk represents the preset scaling factor; output VT t *=a t VT t ;

[0027] The fully connected layer consists of 3 layers of neurons. The first layer input contains 128 neurons, the second layer hidden layer contains 64 neurons; the third layer output contains Q 2 neurons; the hidden layer activation function uses ReLU, the output layer activation function uses Softmax; the output cluster control weight matrix W t ;

[0028] The device execution layer network of the two-layer LSTM network model receives the production device-level state matrix VI t ={VI i t |i∈[1,I]}, where VI i t Represents the production parameters of production equipment i; VI i t =[V i , I i , η i ] t , where V i Indicates the voltage data of production equipment i, I i represents the current data of production equipment i, η i Represents the load rate of production equipment i; through the tensor product VI t * =W t ⊙VI t Perform cross-level feature fusion to output the production equipment control parameter matrix P t , where ⊙ represents the tensor product.

[0029] The upper layer in the device execution layer is used for cluster feature processing. The input is the tensor product of the production device-level state matrix and the inter-cluster control weight matrix. The network structure is a single-layer LSTM with 32 hidden units. The output is the cluster-level control feature h t cluster ;

[0030] The lower layer in the device execution layer is used for device feature processing. The input is the production device-level state matrix. The network structure is a bidirectional LSTM with 64 hidden units in each direction. The output is the device-level dynamic feature h t device ;

[0031] Residual connection: h t final =ReLU(h t cluster ⊕h t device ); where h t final Represents the final residual, ReLU() represents the activation function; ⊕ represents the vector concatenation operation;

[0032] Fully connected mapping Pt=Wp×h t final +bp; where Wp represents the weight matrix of the fully connected layer, and bp represents the bias term of the fully connected layer;

[0033] This application quantifies the interaction between devices through dynamic impact radius and closeness indicators to avoid the subjectivity of artificial cluster division; spectral clustering combined with similarity matrix automatically identifies the functional / spatial correlation of devices to form the optimal cluster division, laying the foundation for hierarchical regulation.

[0034] According to the above solution, S3 contains the following:

[0035] S301: Obtain real-time monitoring sensor data stream and construct a spatiotemporal aligned monitoring data feature set F t ={f t r |r∈[1, R], r is a positive integer}; R is the number of sensor categories; f t r represents the time-frequency domain fusion feature vector of the r-th sensor, f t r =[μ r , ρ r , skewness (S r t-N+1：t )]∪{FFT(S r t-N+1：t )[1:G]}; S r t-N+1：t represents the monitoring data of the rth sensor in the sliding window from t-N+1 to t; N represents the preset sliding window size; μ r represents the mean value in the sliding window of the r-th sensor, ρ r The standard deviation of the sliding window of the rth sensor, skewness() represents the skewness function, and FFT()[1:G] represents the first G main frequency components of the fast Fourier transform;

[0036] S302: Input the monitoring data feature set into the encoder network to generate a hidden state vector h t =LSTM enc (F t-T+1：t ), where F t-T+1：t represents the monitoring data feature set from t-T+1 to t; the decoder network reconstructs the input data F t * =LSTM dec (h t ), the reconstructed time-frequency domain fusion feature vector of the r-th sensor is recorded as f t r* , calculate the abnormal confidence AC at time tt :

[0037] ;

[0038] Among them, α r t represents the attention weight of the r-th sensor, α r t =Softmax(MLP(h t r ), where MLP is a two-layer fully connected network; the input is the encoder network generating the hidden state vector, h t r represents the hidden state of the r-th sensor;

[0039] Encoder 3-layer BiLSTM network, hidden state dimension 64, output hidden state vector h t ;The decoder is a 2-layer LSTM network that reconstructs the input data;

[0040] S303: When the abnormal confidence AC t Greater than the confidence threshold θ t The alarm is triggered when θ t =μ AC +B1×ρ r ×(1+Sigmoid(B2(t-t0))), where μ AC represents the mean of abnormal confidence, ρ r represents the abnormal confidence standard deviation, B1 and B2 represent confidence coefficients, which are preset constants; t0 represents the last maintenance timestamp;

[0041] The data subset F will be transmitted t 'Initialize to an empty set, candidate set C=F t ; For each candidate feature f belonging to the candidate set t r , calculate the information gain IG (f t r ), IG(f t r ) = H (F t ')-H(F t '∪{f t r}); where H(F t ') represents the current transmission data subset F t 'The joint entropy, H(F t '∪{f t r}) means that the feature f t rThe new joint entropy after adding the subset; select the feature with the largest information gain and merge the feature with the largest information gain entropy into the data subset F t ', remove the feature with the largest information gain entropy from the candidate set until |H(F t ')-H(F t )| / H(F t ) is less than the entropy constraint threshold, and the final transmission data subset is output; the entropy constraint threshold is a preset constant; wherein the joint entropy H (F t ) is calculated by the k-nearest neighbor probability estimation method:

[0042] ;

[0043] Among them, b represents the number of equal-width bins, L r ∈{1,2,...,b} represents the box number to which the r-th sensor belongs at time t; p*(L1,...,L R ) means the sensor is in (L1,…,L R ) is the probability estimate of the state.

[0044] This application integrates mean, standard deviation, skewness and FFT main frequency components to comprehensively characterize the equipment operating status and improve the sensitivity of anomaly detection; uses reconstruction error to calculate anomaly confidence, adapts to nonlinear patterns under complex working conditions, and reduces false alarms and missed alarms; the confidence threshold integrates historical statistics and maintenance cycles to adaptively adjust the alarm sensitivity and avoid excessive maintenance; and uses information gain-based data screening to maximize the entropy of transmitted data and reduce redundant data transmission.

[0045] According to the above scheme, S4 includes the following contents:

[0046] Dynamically adjust the hidden layer dimension H of the two-layer LSTM network model based on the transmission data subset. new :

[0047] H new =H base +⌈δ×AC t ⌉;

[0048] Among them, H base represents the original hidden layer dimension, ⌈⌉ represents the rounding operation; δ represents the sensitivity coefficient, and the sensitivity is a system preset constant;

[0049] The newly added neuron weights are initialized with Gaussian distribution and then concatenated with the original neuron weights. The variance of the newly added neuron weights is equal to two divided by the sum of the original hidden layer dimension and the updated hidden layer dimension. The weights of the original neurons are kept unchanged, and only the newly added neuron weights are trained.

[0050] Another aspect of the present application provides a production line energy consumption data management system based on deep learning, which is applied to the above-mentioned production line energy consumption data management method based on deep learning, characterized in that the system includes a distributed fault data acquisition module, an equipment clustering control module, an abnormality monitoring module and a dynamic adjustment module;

[0051] The distributed fault data acquisition module is used to collect and calibrate the three-dimensional positioning coordinates of production equipment, extract the extended time window corresponding to historical fault events, and integrate the spatial distribution of the production equipment position and operating status before and after the fault;

[0052] The equipment clustering control module is used to analyze the spatial distribution density index between production equipment when a production equipment failure occurs, filter and process the historical data to be regulated, and establish a pre-control model for the production equipment based on the filtered data;

[0053] The abnormality monitoring module is used to monitor the operation of the production equipment in the current space, issue an alarm when abnormal operation of the production equipment is detected, and selectively transmit the monitoring data in the current space;

[0054] The dynamic adjustment module is used to retrain the pre-regulation model based on the selectively transmitted data set and dynamically adjust the hidden layer dimension of the pre-regulation model.

[0055] According to the above solution, the distributed fault data acquisition module includes a data acquisition unit and a data fusion unit;

[0056] The data acquisition unit is used to obtain real-time positioning data of each production equipment, generate original coordinates with timestamps, apply Kalman filtering to the original coordinates to generate a standardized coordinate set; extract fault event data from the historical database and generate a fault event set;

[0057] The data fusion unit is used to retrieve the coordinates and distribution status of corresponding production equipment within the event retrieval time interval for the fault event, and generate a fault event set with spatial labels.

[0058] According to the above solution, the device clustering control module includes a similarity analysis unit and a model control unit;

[0059] The similarity analysis unit is used to analyze the neighborhood production equipment set of the production equipment, analyze the closeness index of the neighborhood production equipment set, and cluster the production equipment according to the similarity between them;

[0060] The model control unit is used to output the inter-cluster control weight matrix through the temporal attention mechanism according to the cluster state; and to perform cross-level feature fusion of the production equipment level state and the inter-cluster control weight matrix through tensor product to output the production equipment control parameter matrix.

[0061] According to the above solution, the anomaly monitoring module includes an anomaly confidence analysis unit and a transmission data selection unit;

[0062] The anomaly confidence analysis unit is used to obtain real-time monitoring sensor data streams, construct a spatiotemporally aligned monitoring data feature set, perform data reconstruction based on the monitoring data feature set, and analyze anomaly confidence;

[0063] The transmission data selection unit is used to calculate the information gain of each candidate feature belonging to the candidate set, select the feature with the largest information gain, and incorporate the feature with the largest information gain entropy into the data subset; until the constraint conditions are met, the final transmission data subset is output.

[0064] According to the above scheme, the dynamic adjustment module is used to dynamically adjust the updated hidden layer dimension of the execution layer network of the pre-control model device based on the transmission data subset; the newly added neuron weights are initialized with the Gaussian distribution and then spliced ​​with the original neuron weights; the variance of the newly added neuron weights is equal to two divided by the sum of the original hidden layer dimension and the updated hidden layer dimension; the weights of the original neurons are kept unchanged, and only the newly added neuron weights are trained.

[0065] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention eliminates sensor noise by standardizing coordinates through the Mann filter, thereby improving the spatiotemporal consistency of device location data; dynamically expands the fault event time window to cover the fault incubation period and recovery period, avoids data omissions, and improves the accuracy of subsequent analysis; the present invention quantifies the interaction between devices through dynamic impact radius and tightness index, avoiding the subjectivity of artificial cluster division; spectral clustering is combined with similarity matrix to automatically identify device function / spatial correlation, form optimal cluster division, and lay the foundation for hierarchical control; the present invention integrates mean, standard deviation, skewness and FFT main frequency component to comprehensively represent the data. The system can detect the operating status of equipment and improve the sensitivity of anomaly detection; use reconstruction error to calculate anomaly confidence, adapt to nonlinear patterns under complex working conditions, and reduce false alarms and missed alarms; the confidence threshold integrates historical statistics and maintenance cycles, adaptively adjusts alarm sensitivity, and avoids excessive maintenance; data screening based on information gain maximizes the entropy value of transmitted data and reduces redundant data transmission; elastically adjusts model capacity according to anomaly confidence, enhances model learning ability in high anomaly scenarios, and avoids overfitting or underfitting; adds neuron Gaussian initialization to retain the stability of the original knowledge, and only trains the new part to achieve "hot plug" model upgrade and shorten retraining time. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0067] Figure 1 This is a flow chart of a production line energy consumption data management method based on deep learning according to the present invention;

[0068] Figure 2 This is a structural diagram of a production line energy consumption data management system based on deep learning in the present invention. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0070] See also Figure 1 The present invention provides a technical solution: a production line energy consumption data management method based on deep learning, the method comprising the following steps:

[0071] S1: Collect and calibrate the three-dimensional positioning coordinates of production equipment, extract the extended time window corresponding to historical fault events, and integrate the spatial distribution of the production equipment location and operating status before and after the fault;

[0072] S2: Analyze the spatial distribution density index between production equipment when a production equipment failure occurs, filter and process the historical data to be regulated, and establish a pre-regulation model for the production equipment based on the filtered data;

[0073] S3: Monitor the operation of production equipment in the current space, issue an alarm when abnormal operation of production equipment is detected, and selectively transmit the monitoring data in the current space;

[0074] S4: Retrain the pre-conditioning model based on the selectively transmitted dataset and dynamically adjust the hidden layer dimensions of the pre-conditioning model.

[0075] In S1, the following are included:

[0076] S101: Obtain the real-time positioning data of each production equipment, generate the original coordinates with timestamps, apply Kalman filtering to the original coordinates to generate the standardized coordinate set D = {(x i ,y i , z i , t i )|i∈[1,I], i is a positive integer}; where (x i ,y i , z i ) represents the three-dimensional coordinates of the i-th production equipment, t irepresents the timestamp corresponding to the three-dimensional coordinates of the i-th production equipment, and I represents the total number of production equipment in the space to be analyzed;

[0077] S102: Extract fault event data from the historical database and generate a fault event set E={e m |m∈[1, M], where m is a positive integer}; each fault event contains the faulty production equipment number, the set of affected production equipment, the set of control parameters, and the time interval of the fault event; M represents the total number of fault events;

[0078] S103: The fault event time interval [t m start ,t m end ], mapped to the time axis of the standardized coordinate set D, based on the preset event window △t, the fault event time interval is expanded to generate the event retrieval time interval, recorded as T m ={t|t∈[t m start -△t,t m end +△t]}, t represents the time; for the fault event e m In the event retrieval time interval T m The coordinates and distribution status of the corresponding production equipment are retrieved and recorded as the fault event set E'={(e m , U m )|m∈[1,M]}; where U m Indicates fault event e m The production equipment location distribution status, U m ={(x i ,y i , z i ,t,S （i，m） )|i∈[1,I], i is a positive integer}; where S （i，m） Indicates fault event e m The operating status code of the production equipment i when the event occurs; the operating status code includes the real-time load rate and working mode, and the working mode includes on and off.

[0079] S2 contains the following:

[0080] Example 1: S201: Analyze the distance between each production equipment when the fault occurs based on the fault event set with spatial labels; classify the production equipment whose distance from production equipment i is less than the dynamic influence radius of production equipment i into the neighborhood production equipment set of production equipment i, recorded as N i The dynamic influence radius is equal to the product of the real-time load rate and the influence coefficient, where the influence coefficient is a system preset constant.

[0081] Example 2: S201: Analyze the distance between each production equipment when the fault occurs based on the fault event set with spatial labels; classify the production equipment whose distance from production equipment i is less than the dynamic influence radius of production equipment i into the neighborhood production equipment set of production equipment i, recorded as N i ; Dynamic influence radius R of production equipment i at time t i (t)=R base (1+tanh((VIB i (t)-VIB avg ) / VIB std )); where R base Indicates the equipment foundation influence radius; the equipment foundation influence radius is preset by the system; VIB i (t) represents the vibration frequency of the i-th production equipment at time t; VIB avg Indicates the historical average vibration frequency of the equipment; VIB std Indicates the standard deviation of the equipment's historical vibration frequency;

[0082] Analyze the neighborhood production equipment set N i The closeness index COM i :

[0083] ;

[0084] Among them, n represents the serial number of the production equipment in the neighborhood production equipment set, d （i，n） represents the distance between production equipment n and production equipment i in the set of neighboring production equipment of production equipment i, σ represents the spatial coefficient; exp() represents an exponential function with natural numbers as the base; the spatial coefficient is equal to the product of the spatial attenuation coefficient and the average spacing, the spatial attenuation coefficient is a system preset constant, and the average spacing is the average value of the spacing between production equipment in the monitoring space; analyze the compactness index of all production equipment in the monitoring space and perform normalization processing;

[0085] S202: Analyze the similarity between any two production equipment and construct a similarity matrix A; record the similarity between production equipment i and production equipment j in the monitoring space as A （i，j） ;

[0086] ;

[0087] Among them, d （i，j） It represents the distance between production equipment i and production equipment j in the monitoring space;

[0088] Based on the similarity matrix A, the degree matrix D is constructed. The degree matrix D is a diagonal matrix. （j，j） =∑ i A （i，j）Based on the similarity matrix A and the degree matrix D, a symmetric normalized Laplace matrix is ​​generated. The eigenvector of the Laplace matrix is ​​solved based on the spectral clustering algorithm. The optimal number of clusters is determined by the gap statistics method. The production equipment cluster division result {C1, ..., C Q}; where Q represents the total number of production equipment clusters;

[0089] S203: Construct a pre-control model, which is a two-layer LSTM network model including a cluster coordination layer and a device execution layer, wherein the cluster coordination layer network of the two-layer LSTM network model receives the cluster state matrix VT t ={VT q t |q∈[1,Q], q is a positive integer} as input; VT q t Represents the production equipment cluster C q Vibration temperature fusion characteristics; VT q t =avg[COM k , VIB k , TEM k ] t , k∈C q Among them, VIB k represents the vibration frequency of the kth production equipment, TEM k Represents the temperature of the k-th production equipment, avg[] represents the average function; the inter-cluster control weight matrix W is output through the time attention mechanism t ;

[0090] The device execution layer network of the two-layer LSTM network model receives the production device-level state matrix VI t ={VI i t |i∈[1,I]}, where VI i t Represents the production parameters of production equipment i; VI i t =[V i , I i , η i ] t , where V i Indicates the voltage data of production equipment i, I i represents the current data of production equipment i, η i Represents the load rate of production equipment i; through the tensor product VI t * =W t ⊙VI t Perform cross-level feature fusion to output the production equipment control parameter matrix P t .

[0091] In S3, it contains the following:

[0092] S301: Obtain real-time monitoring sensor data stream and construct a spatiotemporal aligned monitoring data feature set F t ={f t r |r∈[1, R], r is a positive integer}; R is the number of sensor categories; f t r represents the time-frequency domain fusion feature vector of the r-th sensor, f t r =[μ r ,ρ r , skewness (S r t-N+1：t )]∪{FFT(S r t-N+1：t )[1:G]}; S r t-N+1：t represents the monitoring data of the rth sensor in the sliding window from t-N+1 to t; N represents the preset sliding window size; μ r represents the mean value in the sliding window of the r-th sensor, ρ r The standard deviation of the sliding window of the rth sensor, skewness() represents the skewness function, and FFT()[1:G] represents the first G main frequency components of the fast Fourier transform;

[0093] S302: Input the monitoring data feature set into the encoder network to generate a hidden state vector h t =LSTM enc (F t-T+1：t ), where F t-T+1：t represents the monitoring data feature set from t-T+1 to t; the decoder network reconstructs the input data F t * =LSTM dec (h t ), the reconstructed time-frequency domain fusion feature vector of the r-th sensor is recorded as f t r* , calculate the abnormal confidence AC at time t t :

[0094] ;

[0095] Among them, α r t represents the attention weight of the r-th sensor, α r t =Softmax(MLP(h t r), where MLP is a two-layer fully connected network; the input is the encoder network generating the hidden state vector, h t r represents the hidden state of the r-th sensor;

[0096] S303: When the abnormal confidence AC t Greater than the confidence threshold θ t The alarm is triggered when θ t =μ AC +B1×ρ r ×(1+Sigmoid(B2(t-t0))), where μ AC represents the mean of abnormal confidence, ρ r Indicates the abnormal confidence standard deviation, B1 and B2 indicate the confidence coefficients, which are preset constants; t0 indicates the last maintenance timestamp;

[0097] The data subset F will be transmitted t 'Initialize to an empty set, candidate set C=F t ; For each candidate feature f belonging to the candidate set t r , calculate the information gain IG (f t r ), IG(f t r ) = H (F t ')-H(F t '∪{f t r}); where H(F t ') represents the current transmission data subset F t 'The joint entropy, H(F t '∪{f t r}) means that the feature f t r The new joint entropy after adding the subset; select the feature with the largest information gain and merge the feature with the largest information gain entropy into the data subset F t ', remove the feature with the largest information gain entropy from the candidate set until |H(F t ')-H(F t )| / H(F t ) is less than the entropy constraint threshold, and the final transmission data subset is output; the entropy constraint threshold is a preset constant; where the joint entropy H (F t ) is calculated by the k-nearest neighbor probability estimation method:

[0098] ;

[0099] Among them, b represents the number of equal-width bins, L r∈{1,2,...,b} represents the box number to which the r-th sensor belongs at time t; p*(L1,...,L R ) means the sensor is in (L1,…,L R ) is the probability estimate of the state.

[0100] In S4, the following contents are included:

[0101] Dynamically adjust the hidden layer dimension H of the two-layer LSTM network model based on the transmission data subset. new :

[0102] H new =H base +⌈δ×AC t ⌉;

[0103] Among them, H base represents the original hidden layer dimension, ⌈⌉ represents the rounding operation; δ represents the sensitivity coefficient, and the sensitivity is a system preset constant;

[0104] The weights of the newly added neurons are initialized with a Gaussian distribution and then concatenated with the weights of the original neurons. The variance of the weights of the newly added neurons is equal to two divided by the sum of the dimensions of the original hidden layer and the updated hidden layer. The weights of the original neurons are kept unchanged, and only the weights of the newly added neurons are trained.

[0105] See also Figure 2 , the present invention provides a technical solution: a production line energy consumption data management system based on deep learning, the system includes a distributed fault data acquisition module, an equipment clustering control module, an abnormality monitoring module and a dynamic adjustment module;

[0106] The distributed fault data acquisition module is used to collect and calibrate the three-dimensional positioning coordinates of production equipment, extract the extended time window corresponding to historical fault events, and integrate the spatial distribution of the production equipment location and operating status before and after the fault;

[0107] The equipment clustering control module is used to analyze the spatial distribution density index between production equipment when a production equipment failure occurs, filter and process the historical data to be regulated, and establish a pre-control model for the production equipment based on the filtered data;

[0108] The abnormality monitoring module is used to monitor the operation of production equipment in the current space, issue an alarm when abnormal operation of production equipment is detected, and selectively transmit the monitoring data in the current space;

[0109] The dynamic adjustment module is used to retrain the pre-regulation model based on the selectively transmitted data set and dynamically adjust the hidden layer dimension of the pre-regulation model.

[0110] The distributed fault data acquisition module includes a data acquisition unit and a data fusion unit;

[0111] The data acquisition unit is used to obtain the real-time positioning data of each production equipment, generate raw coordinates with timestamps, apply Kalman filtering to the raw coordinates to generate a standardized coordinate set; extract fault event data from the historical database and generate a fault event set;

[0112] The data fusion unit is used to retrieve the coordinates and distribution status of the corresponding production equipment within the event retrieval time interval for the fault event, and generate a fault event set with spatial labels.

[0113] The equipment clustering control module includes a similarity analysis unit and a model control unit;

[0114] The similarity analysis unit is used to analyze the neighborhood production equipment set of the production equipment, analyze the closeness index of the neighborhood production equipment set, and cluster the production equipment according to the similarity between them;

[0115] The model control unit is used to output the inter-cluster control weight matrix through the temporal attention mechanism according to the cluster state; the production equipment-level state and the inter-cluster control weight matrix are cross-level feature fused through tensor product to output the production equipment control parameter matrix.

[0116] The anomaly monitoring module includes an anomaly confidence analysis unit and a transmission data selection unit;

[0117] The anomaly confidence analysis unit is used to obtain real-time monitoring sensor data streams, construct a spatiotemporally aligned monitoring data feature set, reconstruct data based on the monitoring data feature set, and analyze anomaly confidence;

[0118] The transmission data selection unit is used to calculate the information gain of each candidate feature belonging to the candidate set, select the feature with the largest information gain, and incorporate the feature with the largest information gain entropy into the data subset; until the constraint conditions are met, the final transmission data subset is output.

[0119] The dynamic adjustment module is used to dynamically adjust the updated hidden layer dimensions of the execution layer network of the pre-control model device based on the transmission data subset; the newly added neuron weights are initialized through Gaussian distribution and spliced ​​with the original neuron weights; the variance of the newly added neuron weights is equal to two divided by the sum of the original hidden layer dimension and the updated hidden layer dimension; the weights of the original neurons are kept unchanged, and only the weights of the newly added neurons are trained.

[0120] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include," "comprise," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or production apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or production apparatus.

[0121] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A production line energy consumption data management method based on deep learning, characterized in that: The method comprises the following steps: S1: Collect and calibrate the three-dimensional positioning coordinates of production equipment, extract the extended time window corresponding to historical fault events, and integrate the spatial distribution of the production equipment location and operating status before and after the fault; S2: Analyze the spatial distribution density index between production equipment when a production equipment failure occurs, filter and process the historical data to be regulated, and establish a pre-regulation model for the production equipment based on the filtered data; S3: Monitor the operation of production equipment in the current space, issue an alarm when abnormal operation of production equipment is detected, and selectively transmit the monitoring data in the current space; S4: Retrain the pre-conditioning model based on the selectively transmitted dataset and dynamically adjust the hidden layer dimensions of the pre-conditioning model; S2 contains the following: S201: Based on the fault event set with spatial labels, analyze the distance between each production equipment when the fault occurs; the production equipment whose distance to production equipment i is less than the dynamic influence radius of production equipment i is classified into the neighborhood production equipment set of production equipment i, recorded as N i ; Analyze the neighborhood production equipment set N i The closeness index COM i ; Analyze and normalize the tightness indicators of all production equipment in the monitoring space; S202: Analyze the similarity between any two production equipment and construct a similarity matrix A; record the similarity between production equipment i and production equipment j in the monitoring space as A （i，j） ; Based on the similarity matrix A, a degree matrix D is constructed. The degree matrix D is a diagonal matrix. （j，j） =∑ i A （i，j） Based on the similarity matrix A and the degree matrix D, a symmetric normalized Laplace matrix is ​​generated. The eigenvector of the Laplace matrix is ​​solved based on the spectral clustering algorithm. The optimal number of clusters is determined by the gap statistics method. The production equipment cluster division result {C1, ..., C Q }; where Q represents the total number of production equipment clusters; S203: Construct a pre-regulation model, which is a two-layer LSTM network model including a cluster coordination layer and a device execution layer, wherein the cluster coordination layer network of the two-layer LSTM network model receives the cluster state matrix VT t ={VT q t |q∈[1,Q], q is a positive integer} as input; VT q t Represents the production equipment cluster C q Vibration-temperature fusion characteristics; VT q t =avg[COM k ,VIB k ,HE HAS k ] t ,k∈C q ; Among them, VIB k represents the vibration frequency of the kth production equipment, TEM k Represents the temperature of the k-th production equipment, avg[] represents the average function; the inter-cluster control weight matrix W is output through the time attention mechanism t ; The device execution layer network of the two-layer LSTM network model receives the production device-level state matrix VI t ={VI i t |i∈[1,I]}, where VI i t Represents the production parameters of production equipment i; VI i t =[V i , I i , η i ] t , where V i Indicates the voltage data of production equipment i, I i represents the current data of production equipment i, η i Represents the load rate of production equipment i; through the tensor product VI t * =W t ⊙VI t Perform cross-level feature fusion to output the production equipment control parameter matrix P t .

2. The method for managing production line energy consumption data based on deep learning according to claim 1, characterized in that: In S1, the following are included: S101: Obtain the real-time positioning data of each production equipment, generate the original coordinates with timestamps, apply Kalman filtering to the original coordinates to generate the standardized coordinate set D = {(x i ,y i , z i , t i )|i∈[1,I], i is a positive integer}; where (x i ,y i , z i ) represents the three-dimensional coordinates of the i-th production equipment, t i represents the timestamp corresponding to the three-dimensional coordinates of the i-th production equipment, and I represents the total number of production equipment in the space to be analyzed; S102: Extract fault event data from the historical database and generate a fault event set E={e m |m∈[1, M], where m is a positive integer}; each fault event contains the faulty production equipment number, the set of affected production equipment, the set of control parameters, and the time interval of the fault event; M represents the total number of fault events; S103: The fault event time interval [t m start ,t m end ], mapped to the time axis of the standardized coordinate set D, based on the preset event window △t, the fault event time interval is expanded to generate the event retrieval time interval, recorded as T m ={t|t∈[t m start -△t,t m end +△t]}, t represents the time; for the fault event e m In the event retrieval time interval T m The coordinates and distribution status of the corresponding production equipment are retrieved and recorded as the fault event set E'={(e m , U m )|m∈[1,M]}; where U m Indicates fault event e m The production equipment location distribution status, U m ={(x i ,y i , z i ,t,S （i，m） )|i∈[1,I], i is a positive integer}; where S （i，m） Indicates fault event e m The operating status code of the production equipment i when it occurs; the operating status code includes the real-time load rate and working mode, and the working mode includes on and off.

3. The method for managing production line energy consumption data based on deep learning according to claim 2, characterized in that: In S3, it contains the following: S301: Obtain real-time monitoring sensor data stream and construct a spatiotemporal aligned monitoring data feature set F t ={f t r |r∈[1, R], r is a positive integer}; R is the number of sensor categories; f t r represents the time-frequency domain fusion feature vector of the r-th sensor, f t r =[μ r , ρ r , skewness (S r t-N+1：t )]∪{FFT(S r t-N+1：t )[1:G]}; S r t-N+1：t represents the monitoring data of the rth sensor in the sliding window from t-N+1 to t; N represents the preset sliding window size; μ r represents the mean value in the sliding window of the r-th sensor, ρ r The standard deviation of the sliding window of the rth sensor, skewness() represents the skewness function, and FFT()[1:G] represents the first G main frequency components of the fast Fourier transform; S302: Input the monitoring data feature set into the encoder network to generate a hidden state vector h t =LSTM enc (F t-T+1：t ), where F t-T+1：t represents the monitoring data feature set from t-T+1 to t; The decoder network reconstructs the input data F t * =LSTM dec (h t ), the reconstructed time-frequency domain fusion feature vector of the r-th sensor is recorded as f t r* , calculate the abnormal confidence AC at time t t : S303: When the abnormal confidence AC t Greater than the confidence threshold θ t The alarm is triggered when θ t =μ AC +B1×ρ r ×(1+Sigmoid(B2(t-t0))), where μ AC represents the mean of abnormal confidence, ρ r represents the abnormal confidence standard deviation, B1 and B2 represent confidence coefficients, which are preset constants; t0 represents the last maintenance timestamp; The data subset F will be transmitted t 'Initialize to an empty set, candidate set C=F t ; For each candidate feature f belonging to the candidate set t r , calculate the information gain IG (f t r ), select the feature with the largest information gain, and merge the feature with the largest information gain entropy into the data subset F t ', remove the feature with the largest information gain entropy from the candidate set until |H(F t ')-H(F t )| / H(F t ) is less than the entropy constraint threshold, and the final transmission data subset is output; the entropy constraint threshold is a preset constant; wherein the joint entropy H (F t ) is calculated using the k-nearest neighbor probability estimation method.

4. The method for managing production line energy consumption data based on deep learning according to claim 3, characterized in that: In S4, the following contents are included: Dynamically adjust the hidden layer dimension H of the two-layer LSTM network model based on the transmission data subset. new : H new =H base +⌈δ×AC t ⌉; Among them, H base represents the original hidden layer dimension, ⌈⌉ represents the rounding operation; δ represents the sensitivity coefficient, and the sensitivity is a system preset constant; The newly added neuron weights are initialized with Gaussian distribution and then concatenated with the original neuron weights. The variance of the newly added neuron weights is equal to two divided by the sum of the original hidden layer dimension and the updated hidden layer dimension. The weights of the original neurons are kept unchanged, and only the newly added neuron weights are trained.

5. A production line energy consumption data management system based on deep learning, wherein the system is applied to the production line energy consumption data management method based on deep learning according to any one of claims 1 to 4, characterized in that: The system includes a distributed fault data acquisition module, an equipment clustering control module, an abnormality monitoring module and a dynamic adjustment module; The distributed fault data acquisition module is used to collect and calibrate the three-dimensional positioning coordinates of production equipment, extract the extended time window corresponding to historical fault events, and integrate the spatial distribution of the production equipment position and operating status before and after the fault; The equipment clustering control module is used to analyze the spatial distribution density index between production equipment when a production equipment failure occurs, filter and process the historical data to be regulated, and establish a pre-control model for the production equipment based on the filtered data; The abnormality monitoring module is used to monitor the operation of the production equipment in the current space, issue an alarm when abnormal operation of the production equipment is detected, and selectively transmit the monitoring data in the current space; The dynamic adjustment module is used to retrain the pre-regulation model based on the selectively transmitted data set and dynamically adjust the hidden layer dimension of the pre-regulation model.

6. The production line energy consumption data management system based on deep learning according to claim 5, characterized in that: The distributed fault data acquisition module includes a data acquisition unit and a data fusion unit; The data acquisition unit is used to obtain real-time positioning data of each production equipment, generate original coordinates with timestamps, apply Kalman filtering to the original coordinates to generate a standardized coordinate set; extract fault event data from the historical database and generate a fault event set; The data fusion unit is used to retrieve the coordinates and distribution status of corresponding production equipment within the event retrieval time interval for the fault event, and generate a fault event set with spatial labels.

7. The production line energy consumption data management system based on deep learning according to claim 5, characterized in that: The device clustering control module includes a similarity analysis unit and a model control unit; The similarity analysis unit is used to analyze the neighborhood production equipment set of the production equipment, analyze the closeness index of the neighborhood production equipment set, and cluster the production equipment according to the similarity between them; The model control unit is used to output the inter-cluster control weight matrix through the temporal attention mechanism according to the cluster state; and to perform cross-level feature fusion of the production equipment level state and the inter-cluster control weight matrix through tensor product to output the production equipment control parameter matrix.

8. The production line energy consumption data management system based on deep learning according to claim 5, characterized in that: The anomaly monitoring module includes an anomaly confidence analysis unit and a transmission data selection unit; The anomaly confidence analysis unit is used to obtain real-time monitoring sensor data streams, construct a spatiotemporally aligned monitoring data feature set, perform data reconstruction based on the monitoring data feature set, and analyze anomaly confidence; The transmission data selection unit is used to calculate the information gain of each candidate feature belonging to the candidate set, select the feature with the largest information gain, and incorporate the feature with the largest information gain entropy into the data subset; Until the constraints are met, the final transmission data subset is output.

9. The production line energy consumption data management system based on deep learning according to claim 5, characterized in that: The dynamic adjustment module is used to dynamically adjust the updated hidden layer dimension of the execution layer network of the pre-control model device based on the transmission data subset; the newly added neuron weights are initialized with the original neuron weights through Gaussian distribution; the variance of the newly added neuron weights is equal to two divided by the sum of the original hidden layer dimension and the updated hidden layer dimension; the weights of the original neurons are kept unchanged, and only the newly added neuron weights are trained.

Images