Operation monitoring method and system of intelligent weak current system

Abstract

The invention discloses an operation monitoring method and system for an intelligent weak current system, and the method comprises the steps: receiving key electrical parameters, collected by a data collection device in real time, of each target device in the intelligent weak current system, and obtaining a key electrical parameter set; performing feature extraction on each parameter of the key electrical parameter set to obtain an extracted key electrical parameter feature set; the key electrical parameter feature set comprises a current feature, a voltage feature, a temperature feature, a sound parameter feature and a vibration analysis parameter feature; screening suspicious fault features from the key electrical parameter feature set by combining an ant colony algorithm and a simulated annealing algorithm; performing feature fusion on the screened suspicious fault features to obtain fused suspicious fault features; and analyzing the fused suspicious fault features through a trained deep learning model to obtain a fault occurrence type of the intelligent weak current system. According to the invention, the fault detection accuracy of the intelligent weak current system can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring, and in particular to a method and system for monitoring operation of an intelligent weak current system. BACKGROUND

[0002] An intelligent weak current system is an indispensable part of modern buildings, covering building automation, communication networks, security monitoring, and other aspects. It achieves effective management and control of various mechanical and electrical equipment in buildings by comprehensively using advanced information technology, sensing technology, and automatic control technology. The main goal of the intelligent weak current system is to improve the safety, comfort, and energy efficiency of buildings while reducing operating costs.

[0003] In the field of monitoring operation of an intelligent weak current system, existing technical solutions usually adopt a single fault detection method, such as a threshold-based monitoring method or a simple statistical analysis method to identify potential faults. These methods are simple, but their effectiveness is limited when facing complex and variable fault patterns. For example, some threshold-based methods may ignore subtle but important signs of failure, while simple statistical analysis is difficult to capture the relevance and complexity between different electrical parameters.

[0004] Therefore, the existing methods for monitoring operation of an intelligent weak current system are difficult to effectively identify potential faults in complex and variable fault conditions, resulting in low accuracy of fault detection in monitoring operation of an intelligent weak current system. SUMMARY

[0005] The embodiments of the present application provide a method and system for monitoring operation of an intelligent weak current system, which combines ant colony algorithm and simulated annealing algorithm for feature selection, and uses a deep learning model to diagnose suspicious fault features, thereby improving the accuracy and efficiency of fault detection in monitoring operation of an intelligent weak current system.

[0006] An embodiment of the present application provides a method for monitoring operation of an intelligent weak current system, applied to a cloud server, comprising:

[0007] Receiving key electrical parameters of each target device in the intelligent weak current system collected by a data acquisition device in real time to obtain a set of key electrical parameters; the key electrical parameters include current, voltage, temperature, sound parameters, and vibration analysis parameters;

[0008] Performing feature extraction on each parameter in the set of key electrical parameters to obtain a set of extracted key electrical parameter features; the set of key electrical parameter features includes current features, voltage features, temperature features, sound parameter features, and vibration analysis parameter features;

[0009] screening suspicious fault features from the key electrical parameter feature set by combining an ant colony algorithm and a simulated annealing algorithm;

[0010] performing feature fusion on the screened suspicious fault features to obtain fused suspicious fault features;

[0011] analyzing the fused suspicious fault features by using the trained deep learning model to obtain a fault occurrence type of the intelligent weak current system.

[0012] As an improvement of the above scheme, the screening suspicious fault features from the key electrical parameter feature set by combining an ant colony algorithm and a simulated annealing algorithm comprises:

[0013] performing ant colony algorithm parameter initialization and simulated annealing algorithm parameter initialization;

[0014] randomly selecting a part of key electrical parameter features from the key electrical parameter feature set as an initial solution;

[0015] each ant selects a key electrical parameter feature according to a current pheromone concentration τ ij and a heuristic information η ij and uses a probability formula p ij to select the key electrical parameter feature, construct a feature subset, and construct all the feature subsets to form a solution space of the current iteration; the higher the pheromone concentration, the greater the probability of selection of the related key electrical parameter feature, and the more relevant the key electrical parameter feature to fault diagnosis;

[0016] randomly selecting a feature subset from the current solution space as a current solution;

[0017] generating a new feature subset by randomly adding or deleting a key electrical parameter feature in the current solution, and taking the new feature subset as a new solution;

[0018] evaluating an evaluation function value E' of the new solution and comparing it with the current solution E; if E' < E, the new solution is accepted; if E' ≥ E, it is determined according to a preset acceptance probability formula whether to accept the new solution; the evaluation function is used to evaluate the effectiveness of different feature subsets, and the higher the evaluation function value, the more likely the feature subset contains suspicious fault features;

[0019] after each iteration, the ants locally update the pheromone according to the path quality of the path selected by them using an improved local update formula;

[0020] at the end of each iteration, the pheromone is updated according to the selection results of all ants and through an improved global update formula to strengthen the high-quality path;

[0021] Heuristic information is adjusted based on the feature performance during the iteration process to guide ants to prefer high-quality key electrical parameter features. The higher the heuristic information, the higher the probability that the relevant key electrical parameter feature will be selected as a suspected fault feature.

[0022] Adjusted the adaptive cooling coefficient in the simulated annealing algorithm to better balance the relationship between global search and local optimization;

[0023] When a preset iteration stop condition is met, the algorithm ends and outputs a final feature combination, thereby obtaining a suspected fault feature from the key electrical parameter feature set.

[0024] As an improvement to the above solution, the probability formula p ij for: p i is the probability that ant k selects feature i; τ i is the pheromone concentration of feature i; η i is the heuristic information of feature i; α and β are preset algorithm parameters that control the influence of pheromone and heuristic information respectively;

[0025] The acceptance probability formula is: P(e, e′, T) is the probability of accepting a new solution; E is the evaluation function value of the current solution; E′ is the evaluation function value of the candidate solution; T is the current temperature;

[0026] After each iteration, the ant locally updates the pheromone according to the path quality of the path it chooses. The improved local update formula is: is the pheromone concentration of feature i at time t; is the pheromone concentration of feature i at time t+1; ρ(t) is the pheromone volatilization rate at time t; is the contribution of ant k to feature i at time t; the contribution formula is: Q is a constant representing the pheromone intensity, which is used to adjust the scale of pheromone update; L best is the length of the current best solution; d i is the importance score of feature i; D max is the highest importance score among all features; ρ0 is the initial volatility; ρ max is the maximum volatility; L best is the path length of the current best solution; L avg is the path length of the current average solution; γ is the parameter that controls the rate of change of volatility;

[0027] The improved global update formula is: is the pheromone concentration of feature i at time t; is the pheromone concentration of feature i at time t+1; ρ(t) is the pheromone volatility rate at time t; m is the number of ants; is the weighted contribution of ant k to feature i at time t; the weighted contribution formula is: L k is the path length of ant k; Q is a constant representing the pheromone intensity, which is used to adjust the scale of pheromone update; d i is the importance score of feature i; D max is the highest importance score among all features; w k is the weighting factor of ant k, which can be adjusted according to the path quality of ant k; β is a parameter that controls the speed of change of the weighting factor;

[0028] The dynamic heuristic information update formula used to adjust the heuristic information according to the feature performance in the iterative process is: is the dynamic heuristic information of feature i in time; is the initial heuristic information of feature i; γ is the dynamic adjustment coefficient, which is used to control the dynamic change degree of heuristic information; is the importance score of feature i at time t; f max is the highest importance score among all features; represents the importance score of feature i calculated using the random forest model M;

[0029] The adjustment formula of the adaptive cooling coefficient is: α(t) is the cooling coefficient of the tth iteration; α init is the preset initial cooling coefficient; β is the cooling coefficient decay rate, which determines the speed at which the cooling coefficient decreases with the increase of the number of iterations; t is the current number of iterations; MaxIter is the maximum number of iterations.

[0030] As an improvement to the above solution, the feature fusion of the screened suspicious fault features to obtain the fused suspicious fault features includes:

[0031] The screened suspicious fault features are input into the feature fusion algorithm formula for feature fusion to obtain the fused suspicious fault features; the feature fusion algorithm formula is:

[0032] F fusion is the fused feature vector; f i is the original feature vector of the i-th suspected fault feature; g i =σ(W g ·f i +b g) is the vector of the ith suspicious fault feature after nonlinear transformation; W g and b g are the weight and bias of the nonlinear transformation layer; σ is the ReLU or Sigmoid activation function; score(g i ) = W s · σ(W g · g i +b g )+b s is the score of the ith suspicious fault feature, used to calculate the adaptive weight; W s and b s are the weight and bias of the score calculation layer; exp is the exponential function used to calculate the attention weight; j represents the index of the feature vector, for traversing the feature vector of each suspicious fault feature, calculating the sum of the exponential function values of the scores of each property after nonlinear transformation, thereby calculating the weight of the feature vector of each suspicious fault feature, ensuring that the sum of the weights is 1; N represents the total number of feature vectors of suspicious fault features.

[0033] As an improvement of the above scheme, the above-mentioned scheme further comprises:

[0034] inputting the suspicious fault feature into the input layer of the deep learning model of the hybrid model comprising multi-scale convolution, attention mechanism, GRU and adaptive learning rate, and sequentially performing local feature multi-scale extraction by the multi-scale convolution layer of the deep learning model, feature dependency relationship learning by the attention layer, long-term dependency relationship capture by the GRU layer and classification by the full connection layer, and then outputting the fault occurrence type of the intelligent weak current system by the output layer of the trained deep learning model;

[0035] The deep learning model comprises:

[0036] an input layer for receiving the preprocessed fusion feature data;

[0037] a multi-scale convolution layer for extracting multi-scale local features; the multi-scale convolution layer uses a plurality of filters wherein K k is the size of the kth filter, and F is the number of filters; the convolution operation is represented as: h (k) = f(W (k) *x+b (k) ), * represents convolution operation; b (k) is a bias vector; the convolution outputs h (k) of different scales are spliced to obtain the final convolution output H;

[0038] An attention layer for learning inter-dependencies between features; the attention weights of the attention layer are represented as: a = softamx(U · f(V · H + c) + d), U, V, c, d are weight and bias parameters; a is the attention weight vector; H is the convolution output;

[0039] A GRU layer for processing sequential data and capturing long-term dependencies; the GRU layer uses a Gated Recurrent Unit to process sequential data, represented as:

[0040] z t = σ(U z h t-1 +W z (H t ☉a t )+b z )

[0041] r t = σ(U r h t-1 +W r (H t ☉a t )+b r )

[0042]

[0043] where z t , r t and h t are the update gate, reset gate and hidden state respectively; U z , U r , U h , W z , W r , W h are preset weights, b z , b r , b h are preset bias parameters; σ is the sigmoid activation function; tanh is the hyperbolic tangent function; H t is the output of the convolution layer; a t is the attention weight;

[0044] A fully connected layer for classification tasks; the fully connected layer is: z = f(Uh T +c), U is a weight matrix; c is a bias vector; h T is the GRU hidden state at the last time step;

[0045] An output layer for predicting the output failure type; the output layer uses a softmax function for classification: V is a preset weight matrix; d is a preset bias vector; is a preset prediction probability vector; z is a classification output of the full connection layer.

[0046] Another embodiment of the present application provides an operation monitoring system of an intelligent weak current system, comprising:

[0047] A receiving module is configured to receive key electrical parameters of each target device in the intelligent weak current system collected by a data acquisition device in real time, and obtain a key electrical parameter set; the key electrical parameters include current, voltage, temperature, sound parameters and vibration analysis parameters;

[0048] A feature extraction module is configured to perform feature extraction on each parameter in the key electrical parameter set, and obtain an extracted key electrical parameter feature set; the key electrical parameter feature set includes current features, voltage features, temperature features, sound parameter features and vibration analysis parameter features;

[0049] A feature screening module is configured to screen suspicious fault features from the key electrical parameter feature set by combining an ant colony algorithm and a simulated annealing algorithm;

[0050] A feature fusion module is configured to perform feature fusion on the screened suspicious fault features, and obtain fused suspicious fault features;

[0051] A fault analysis module is configured to analyze the fused suspicious fault features by using a trained deep learning model, and obtain a fault occurrence type of the intelligent weak current system.

[0052] As an improvement of the above scheme, the feature screening module is specifically configured to:

[0053] Perform ant colony algorithm parameter initialization and simulated annealing algorithm parameter initialization;

[0054] Randomly select a part of key electrical parameter features from the key electrical parameter feature set as initial solutions;

[0055] Each ant selects key electrical parameter features according to current pheromone concentration τ ij and heuristic information η ij and uses a probability formula p ij to select key electrical parameter features, construct feature subsets, and construct all the constructed feature subsets into a solution space of the current iteration; the higher the pheromone concentration, the greater the probability of selection of the related key electrical parameter features, and the more relevant the key electrical parameter features to fault diagnosis;

[0056] Randomly select a feature subset from the current solution space as a current solution;

[0057] A new feature subset is generated by randomly adding or deleting one key electrical parameter feature in the current solution, and the new feature subset is taken as a new solution;

[0058] The evaluation function value E' of the new solution is evaluated and compared with the current solution E; if E' < E, the new solution is accepted; if E' ≥ E, it is determined whether to accept the new solution according to a preset acceptance probability formula; the evaluation function is used to evaluate the effectiveness of different feature subsets, and the higher the evaluation function value, the more likely the feature subset contains suspicious fault features;

[0059] After each iteration, the ants locally update the pheromone according to the path quality of the path selected by them using an improved local update formula;

[0060] At the end of each iteration, the pheromone is updated according to the selection results of all ants and through an improved global update formula, and the high-quality path is strengthened;

[0061] The heuristic information is adjusted according to the feature performance in the iteration process, guiding the ants to preferentially select high-quality key electrical parameter features; the higher the heuristic information, the higher the probability of the related key electrical parameter feature being selected as a suspicious fault feature;

[0062] The adaptive cooling coefficient in the simulated annealing algorithm is adjusted to better balance the relationship between global search and local optimization;

[0063] When the preset iteration stopping condition is met, the algorithm ends and outputs the final feature combination, so that the suspicious fault features are obtained from the key electrical parameter feature set.

[0064] As an improvement of the above scheme, the probability formula p ij is: p i is the probability of ant k selecting feature i; τ i is the pheromone concentration of feature i; η i is the heuristic information of feature i; and α and β are preset algorithm parameters, respectively controlling the influence degree of pheromone and heuristic information;

[0065] The acceptance probability formula is: P(e, e', T) is the probability of accepting the new solution; E is the evaluation function value of the current solution; E' is the evaluation function value of the candidate solution; and T is the current temperature;

[0066] After each iteration, the ants locally update the pheromone according to the path quality of the path selected by them using an improved local update formula: is the pheromone concentration of feature i at time t; is the pheromone concentration of feature i at time t + 1; p(t) is the pheromone evaporation rate at time t; m is the number of ants; is the contribution of ant k to feature i at time t; the contribution formula is: Q is a constant representing the pheromone intensity, used to adjust the scale of pheromone update; d best is the length of the current best solution; d i is the importance score of feature i; D max is the highest importance score among all features; w p0 is the initial evaporation rate; p max is the maximum evaporation rate; L best is the path length of the current best solution; L avg is the path length of the current average solution; γ is a parameter controlling the speed of evaporation rate change;

[0067] The improved global update formula is: is the pheromone concentration of feature i at time t + 1; p(t) is the pheromone evaporation rate at time t; m is the number of ants; is the pheromone concentration of feature i at time t + 1; p(t) is the pheromone evaporation rate at time t; m is the number of ants; is the weighted contribution of ant k to feature i at time t; the weighted contribution formula is: L k is the path length of ant k; Q is a constant representing the pheromone intensity, used to adjust the scale of pheromone update; d i is the importance score of feature i; D max is the highest importance score among all features; w k is the weighting factor of ant k, which can be adjusted according to the path quality of ant k; β is a parameter controlling the speed of change of the weighting factor;

[0068] The dynamic heuristic information update formula used in the heuristic information according to the feature performance in the iteration process is: is the dynamic heuristic information of feature i at time t; is the initial heuristic information of feature i; γ is a dynamic adjustment coefficient, used to control the degree of dynamic change of the heuristic information; is the importance score of feature i at time t; f max is the highest importance score among all features; represents the importance score of feature i calculated using the random forest model M;

[0069] The adjustment formula of the adaptive cooling coefficient is: α(t) is the cooling coefficient of the t-th iteration; α initis a preset initial cooling coefficient; β is a cooling coefficient decay rate, which determines the speed of the decrease of the cooling coefficient with the increase of the iteration number; t is the current iteration number; and MaxIter is the maximum iteration number.

[0070] As an improvement of the above scheme, the feature fusion module is specifically configured to:

[0071] input the screened suspicious fault features into a feature fusion algorithm formula for feature fusion to obtain fused suspicious fault features; the feature fusion algorithm formula is:

[0072] F fusion is a fused feature vector; f i is an original feature vector of the i th suspicious fault feature; g i =σ(W g ·f i +b g ) is a vector after nonlinear transformation of the i th suspicious fault feature; W g and b g are weights and biases of the nonlinear transformation layer; σ is a ReLU or Sigmoid activation function; score(g i )=W s ·σ(W g ·g i +b g )+b s is a score of the i th suspicious fault feature, which is used to calculate an adaptive weight; W s and b s are weights and biases of a score calculation layer; exp is an exponential function, which is used to calculate an attention weight; j represents an index of a feature vector, is used to traverse the feature vector of each suspicious fault feature, calculate a sum of exponential function values of scores after nonlinear transformation of each property, and thus calculate a weight of the feature vector of each suspicious fault feature, so as to ensure that the sum of the weights is 1; and N represents a total number of the feature vectors of the suspicious fault features.

[0073] As an improvement of the above scheme, the fault analysis module is specifically configured to:

[0074] input the suspicious fault features into an input layer of a deep learning model of a hybrid model containing multi-scale convolution, attention mechanism, GRU and adaptive learning rate, sequentially perform local feature multi-scale extraction through a multi-scale convolution layer of the deep learning model, feature dependency relationship learning through an attention layer, long-term dependency relationship capture through a GRU layer, and classification through a full connection layer, and then output, through an output layer of the trained deep learning model, a predicted fault occurrence type of the intelligent weak current system.

[0075] The deep learning model comprises:

[0076] an input layer for receiving pre-processed fused feature data;

[0077] a multi-scale convolutional layer for extracting multi-scale local features; the multi-scale convolutional layer uses multiple filters where K k is the size of the k-th filter, and F is the number of filters; the convolution operation is denoted as: h (k) = f(W (k) *x+b (k) ), * denotes the convolution operation; b (k) is a bias vector; the convolution outputs h (k) of different scales are concatenated to obtain the final convolution output H;

[0078] an attention layer for learning the mutual dependency between features; the attention weight of the attention layer is denoted as: a = softmax(U·f(V·H+c)+d), U, V, c, d are weight and bias parameters; a is an attention weight vector; H is the convolution output;

[0079] a GRU layer for processing sequence data and capturing long-term dependencies; the GRU layer uses a gated recurrent unit to process sequence data, denoted as:

[0080] z t = σ(U z h t-1 +W z (H t ☉a t )+b z )

[0081] r t = σ(U r h t-1 +W r (H t ☉a t )+b r )

[0082]

[0083] where z t , r t and h t are the update gate, the reset gate and the hidden state, respectively; U z , U r , U h , W z , W r , W h are preset weights, and b z , b r,b h is the preset bias parameter; σ is the sigmoid activation function; tanh is the hyperbolic tangent function; H t is the output of the convolutional layer; a t is the attention weight;

[0084] The fully connected layer for classification tasks is: z = f(Uh T +c), U is the weight matrix; c is the bias vector; h T is the GRU hidden state at the last time step;

[0085] The output layer is used to predict the output fault type; the output layer uses the softmax function for classification: V is the preset weight matrix; d is the preset bias vector; is the preset prediction probability vector; z is the classification output of the fully connected layer.

[0086] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0087] The embodiment of the present invention collects the key electrical parameters of each target device in the intelligent weak current system in real time through a data acquisition device to form a key electrical parameter set. Subsequently, these parameters are subjected to feature extraction to construct a key electrical parameter feature set. Next, a method combining an ant colony algorithm and a simulated annealing algorithm is used to screen out suspicious fault features from the feature set. The screened suspicious fault features are then subjected to feature fusion to generate fused suspicious fault features. Finally, the fused suspicious fault features are analyzed by a pre-trained deep learning model to accurately determine the fault type of the intelligent weak current system. The embodiment of the present invention significantly improves the accuracy and efficiency of fault detection by combining an ant colony algorithm with a simulated annealing algorithm for feature selection and using a deep learning model to perform fault diagnosis on the fused suspicious fault features. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 This is a flow chart of an operation monitoring method for an intelligent weak current system provided by one embodiment of the present invention;

[0089] Figure 2 The diagram is a structural diagram of an operation monitoring system for an intelligent weak current system provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0090] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all the other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0091] Referring to Figure 1 is a flowchart of an intelligent weak current system operation monitoring method provided by an embodiment of the present application. The intelligent weak current system operation monitoring method is applied to a cloud server and includes steps S10 to S14.

[0092] S10, receiving key electrical parameters of each target device in the intelligent weak current system collected by a data collection device in real time to obtain a key electrical parameter set; the key electrical parameters include current, voltage, temperature, sound parameters and vibration analysis parameters;

[0093] S11, performing feature extraction on each parameter of the key electrical parameter set to obtain an extracted key electrical parameter feature set; the key electrical parameter feature set includes current features, voltage features, temperature features, sound parameter features and vibration analysis parameter features;

[0094] S12, screening suspicious fault features from the key electrical parameter feature set by combining an ant colony algorithm and a simulated annealing algorithm;

[0095] S13, performing feature fusion on the screened suspicious fault features to obtain fused suspicious fault features;

[0096] S14, analyzing the fused suspicious fault features by using a trained deep learning model to obtain a fault occurrence type of the intelligent weak current system.

[0097] In this embodiment, specifically, each target device in the intelligent weak current system can be monitored in real-time by data acquisition devices. These data acquisition devices continuously collect key electrical parameters such as current, voltage, temperature, and possibly some unconventional monitoring parameters such as sound and vibration analysis parameters. These parameters are obtained through sensors installed on each device, for example, current transformers for measuring current, voltmeters for measuring voltage, thermocouples or infrared sensors for measuring temperature, and microphones or acoustic sensors for capturing sound signals, and accelerometers or other vibration sensors for monitoring vibration conditions. All these data are transmitted to the cloud server in real time, forming a set of key electrical parameters for all monitoring points, providing a basis for subsequent data analysis and fault diagnosis. Among them, in the operation monitoring of the intelligent weak current system, the key electrical parameters collected include current, voltage, temperature, sound parameters and vibration analysis parameters, and their selection and collection aim to capture key information under normal operation and abnormal state of the equipment. Specifically, current and voltage are basic indicators for evaluating the health of electrical systems. Abnormal increase of current may indicate overload or short circuit problems, while voltage fluctuations may indicate unstable power supply or poor internal connection. By monitoring the trend of current and voltage, potential faults in the electrical system can be discovered in time. Temperature parameter is another important monitoring indicator, especially for heat-sensitive electrical equipment such as motors and transformers. Excessive temperature may be caused by overload, insulation damage or poor heat dissipation, which may lead to premature aging or even failure of the equipment. Therefore, monitoring temperature can help to early warning of overheating risk and avoid potential electrical accidents. For sound parameters, electrical equipment will produce certain noise when working normally, and when the equipment fails, the noise will often change. By analyzing the sound parameters, the abnormal sound characteristics of the equipment such as abnormal noise, whistling sound, etc. can be captured, which helps to discover potential mechanical or electrical problems early. For vibration analysis parameters, vibration analysis is a very effective fault diagnosis tool, especially for rotating machinery and motors. By monitoring the vibration characteristics of the equipment, mechanical problems such as imbalance and abnormal vibration, and electrical problems such as motor rotor failure can be found. Vibration analysis not only helps to identify the existence of faults, but also provides information on the specific location and severity of the fault.

[0098] In addition, after collecting the key electrical parameters, the next step is to extract features from these parameters. This process involves extracting specific attributes or patterns from the raw data that are helpful for fault detection and classification. For example, for current features, statistical quantities such as peak value, mean value, variance of current waveform can be extracted; for voltage features, maximum value, minimum value, average value and fluctuation range of voltage can be extracted; temperature features can include maximum temperature, minimum temperature, temperature rise rate, etc.; sound parameter features may include sound spectrum characteristics, sound pressure level, etc.; and vibration analysis parameter features may include vibration amplitude, frequency components, etc. Through these specific mathematical methods and signal processing techniques, we can extract a more representative set of key electrical parameter features from a large amount of raw data, which provides high-quality data support for subsequent fault feature screening, feature fusion and fault type analysis. This can improve the accuracy and efficiency of fault diagnosis, helping maintenance personnel to quickly locate and solve problems, ensuring the stable operation of intelligent weak current systems.

[0099] The embodiment of the application collects key electrical parameters of each target device in the intelligent weak current system in real time through the data acquisition device, forming a set of key electrical parameters. Then, feature extraction is performed on these parameters to construct a set of key electrical parameter features. Next, a method combining ant colony algorithm and simulated annealing algorithm is used to select suspicious fault features from the feature set. Then, the selected suspicious fault features are fused to generate fused suspicious fault features. Finally, the fused suspicious fault features are analyzed by a pre-trained deep learning model to accurately determine the fault type of the intelligent weak current system. The embodiment of the application selects features by combining ant colony algorithm and simulated annealing algorithm, and uses a deep learning model to diagnose the fused suspicious fault features, significantly improving the accuracy and efficiency of fault detection.

[0100] As an improvement of the above-mentioned scheme, the method of selecting suspicious fault features from the set of key electrical parameter features by combining ant colony algorithm and simulated annealing algorithm comprises:

[0101] performing ant colony algorithm parameter initialization and simulated annealing algorithm parameter initialization;

[0102] randomly selecting a part of the key electrical parameter features from the set of key electrical parameter features as initial solutions;

[0103] each ant updates the current pheromone concentration τ ij and heuristic information η ij and uses a probability formula p ijselecting key electrical parameter features, constructing feature subsets, and constructing all the constructed feature subsets into a solution space of the current iteration; the higher the pheromone concentration, the greater the probability of selecting the relevant key electrical parameter feature, and the more relevant the key electrical parameter feature to fault diagnosis;

[0104] randomly selecting a feature subset from the current solution space as a current solution;

[0105] generating a new feature subset by randomly adding or deleting a key electrical parameter feature in the current solution, and taking the new feature subset as a new solution;

[0106] evaluating the evaluation function value E' of the new solution and comparing it with the current solution E; if E' < E, the new solution is accepted; if E' ≥ E, it is determined according to a preset acceptance probability formula whether to accept the new solution; the evaluation function is used to evaluate the effectiveness of different feature subsets, and the higher the evaluation function value, the more likely the feature subset contains suspicious fault features;

[0107] After each iteration, the ants locally update the pheromone according to the path quality of the path selected by them using an improved local update formula;

[0108] At the end of each iteration, the pheromone is updated according to the selection results of all ants and through an improved global update formula, and the high-quality path is strengthened;

[0109] According to the feature performance in the iteration process, the heuristic information is adjusted to guide the ants to preferentially select high-quality key electrical parameter features; the higher the heuristic information, the higher the probability of selecting the relevant key electrical parameter feature as a suspicious fault feature;

[0110] The adaptive cooling coefficient in the simulated annealing algorithm is adjusted to better balance the relationship between global search and local optimization;

[0111] When the preset iteration stopping condition is met, the algorithm ends and outputs the final feature combination, so that a suspicious fault feature is obtained from the key electrical parameter feature set.

[0112] In this embodiment, the specific process is as follows:

[0113] 1. Parameter initialization: First, the parameters of the ant colony algorithm are initialized: the number of ants m = 10; the evaporation rate of pheromone rho = 0.5; the heuristic information factor alpha = 1; the pheromone importance factor beta = 5. Next, the parameters of the simulated annealing algorithm are initialized: the initial temperature T0 = 1000; the cooling coefficient cooling_rate = 0.99; the maximum number of iterations max_iter = 100. Then, the initial solution is generated: the key electrical parameter feature set: I: peak current (I1), average current (I2), current harmonic (I3); V: peak voltage (V1), average voltage (V2), voltage harmonic (V3); T: environmental temperature (T1), equipment internal temperature (T2); S: noise intensity (S1), frequency component (S2); A: vibration amplitude (A1), frequency component (A2); a part of the features are randomly selected as the initial solution: for example, the initial solution can be {I1, V1, T1, S1, A1}.

[0114] 2. Pheromone initialization: the pheromone concentration of all features is initialized to the same value, such as 1.0. The heuristic information is initialized: the heuristic information of all features is initialized to the same value, such as 1.0.

[0115] 3. Constructing feature subsets: each ant selects features according to the current pheromone concentration and heuristic information. For example, ant 1 may select {I1, V1, T1, S1, A1} according to the pheromone concentration and heuristic information, ant 2 may select {I2, V2, T2, S2, A2}, etc. The constructed feature subsets constitute the solution space of the current iteration. The influence of pheromone and heuristic information: the higher the pheromone concentration, the greater the probability of a feature being selected as a suspicious fault feature, which means that the feature is more relevant to fault diagnosis. The higher the heuristic information, the relatively higher the probability of a feature being selected as a suspicious fault feature, which indicates that the feature may be more important for fault diagnosis.

[0116] 4. Generating new solutions: a feature subset is randomly selected from the current solution space as the current solution. For example, the feature subset of ant 1 {I1, V1, T1, S1, A1} is selected. A feature is randomly added or deleted in the current solution to generate a new solution. For example, V1 is deleted and V2 is added to obtain the new solution {I1, T1, S1, A1, V2}.

[0117] 5. Evaluation of new solution: Evaluate the value of the evaluation function for the new solution. The evaluation function can be based on the accuracy of the diagnosis of whether the features are suspicious fault features, for example, using cross-validation accuracy. If the evaluation function value of the new solution is better than the current solution, accept the new solution; otherwise, decide whether to accept the new solution according to the acceptance probability formula of the simulated annealing algorithm. For example, the accuracy of the new solution is 90%, while the current solution is 85%. Because the new solution is better, directly accept the new solution.

[0118] 6. Pheromone update: After each iteration, update the pheromone locally based on the quality of the paths selected by the ants. At the end of each iteration, update the pheromone globally based on the selection results of all ants. The update formula takes into account the quality of the path and the persistence of the pheromone. The effect of pheromone update: updating the pheromone concentration can reflect the effectiveness of the features for diagnosing faults, further improving the probability of selecting high-quality features in subsequent iterations.

[0119] 7. Heuristic information adjustment: Adjust the heuristic information according to the performance of the features in the iteration process. Guide the ants to preferentially select high-quality key electrical parameter features (i.e., more likely to be suspicious fault features). The effect of heuristic information adjustment: by adjusting the heuristic information, the importance of features that perform well in the diagnosis process can be further emphasized, thereby improving their selection probability.

[0120] 8. Adjust the cooling coefficient in the simulated annealing algorithm: appropriately adjust the cooling coefficient to maintain a balance between global search and local optimization. As the number of iterations increases, the temperature gradually decreases, reducing the probability of accepting poor solutions. The effect of the cooling coefficient: by adjusting the cooling coefficient, more exploration can be allowed in the early stage, while more attention is paid to local optimal solutions in the later stage to achieve a global optimal solution.

[0121] 9. Stopping condition: when the preset number of iterations or convergence condition is reached, the algorithm stops.

[0122] For example, assume that several iterations have been performed, the following is the detail of the jth iteration: current pheromone concentration: I1=3.0, I2=2.0, I3=1.0; V1=2.5, V2=2.0, V3=1.5; T1=3.5, T2=2.5; S1=3.0, S2=2.0; A1=3.5, A2=2.5. Current heuristic information: I1=1.5, I2=1.0, I3=0.5; V1=1.5, V2=1.0, V3=0.5; T1=2.0, T2=1.5; S1=1.5, S2=1.0; A1=2.0, A2=1.5. Current solution: {I1, V1, T1, S1, A1} (assume this is a solution randomly selected from the solution space). Generate new solution: for example, remove V1 and add V2 -> new solution: {I1, T1, S1, A1, V2}. Evaluate new solution: use cross-validation accuracy as evaluation function. Assume the evaluation function value of the current solution is 85% and the evaluation function value of the new solution is 90%. Because the evaluation function value of the new solution is higher, accept the new solution. Local update pheromone: update the pheromone concentration related to V2. For example, the pheromone concentration of V2 is updated from 1.5 to 2.0. Global update pheromone: update the pheromone concentration of the path selected by all ants. For example, the pheromone concentration of all selected features is updated according to their evaluation function values. Adjust heuristic information: increase the heuristic information weight of V2. For example, the heuristic information of V2 is updated from 0.5 to 1.0. Adjust temperature: T10=1000*0.99^9. Continue iteration until the stopping condition is met. After several rounds of iteration, the final feature combination can be {I1, T1, S1, A1, V2}. These features are considered as the key electrical parameter features that are most likely to contain suspicious fault features.

[0123] In this embodiment, the key electrical parameter features highly correlated with fault diagnosis are screened by integrating the ant colony algorithm and the simulated annealing algorithm. Specifically, first, the algorithm parameter initialization is performed, including setting the number of ants, the evaporation rate of pheromone, the heuristic information factor, the importance factor of pheromone, and the initial temperature, the cooling coefficient and the maximum number of iterations in the simulated annealing algorithm. Subsequently, a part of features is randomly selected from the key electrical parameter feature set as the initial solution. On this basis, each virtual "ant" selects features according to the current pheromone concentration and heuristic information to construct a feature subset. The higher the pheromone concentration, the higher the probability of feature selection, which indicates that the feature is more strongly correlated with fault diagnosis. At the same time, the higher the heuristic information, the relatively higher the probability of feature selection, indicating that the feature is more important for fault diagnosis. Next, in each iteration, the algorithm generates a new solution by randomly adding or deleting a feature in the current solution and evaluates the effectiveness of the new solution. If the evaluation function value of the new solution is higher than that of the current solution, the new solution is accepted; otherwise, whether to accept the new solution is determined according to the acceptance probability formula in the simulated annealing algorithm. After each iteration, the pheromone concentration is updated according to the quality of the path selected by the ant to strengthen the high-quality path, and the heuristic information is adjusted according to the performance of the feature to further improve the probability of selecting high-quality features. In addition, by adjusting the cooling coefficient in the simulated annealing algorithm, the relationship between global search and local optimization can be better balanced to ensure that the algorithm can both extensively explore in the early stage and focus on better solutions in the later stage. In summary, this embodiment gradually improves the accuracy of suspicious fault feature screening through continuous iteration and optimization, so that the final selected feature subset is more likely to contain suspicious fault features. Specifically, through the dynamic adjustment of pheromone concentration and heuristic information, the algorithm can more intelligently select features that perform well in the diagnosis process, thereby improving the efficiency of suspicious fault feature screening. By adjusting the cooling coefficient in the simulated annealing algorithm, the algorithm can find the best solution in the global range while avoiding falling into a local optimal solution. Through the dynamic adjustment mechanism of pheromone concentration, heuristic information and cooling coefficient, the adaptability and robustness of the algorithm to complex problems are improved. Using the improved local and global pheromone update formula, the algorithm can converge to the optimal solution more quickly, thereby effectively screening the most relevant fault features.

[0124] As an improvement of the above scheme, the probability formula p ij is: p i is the probability of ant k selecting feature i; τ i is the pheromone concentration of feature i; η i is the heuristic information of feature i; and α and β are preset algorithm parameters controlling the influence degree of pheromone and heuristic information, respectively;

[0125] The acceptance probability formula is: P(e, e′, T) is the probability of accepting a new solution; E is the evaluation function value of the current solution; E′ is the evaluation function value of the candidate solution; T is the current temperature;

[0126] After each iteration, the ant locally updates the pheromone according to the path quality of the path it chooses. The improved local update formula is: is the pheromone concentration of feature i at time t; is the pheromone concentration of feature i at time t+1; ρ(t) is the pheromone volatilization rate at time t; is the contribution of ant k to feature i at time t; the contribution formula is: Q is a constant representing the pheromone intensity, which is used to adjust the scale of pheromone update; L best is the length of the current best solution; d i is the importance score of feature i; D max is the highest importance score among all features; ρ0 is the initial volatility; p max is the maximum volatility; L best is the path length of the current best solution; L avg is the path length of the current average solution; γ is the parameter that controls the rate of change of the volatility;

[0127] The improved global update formula is: is the pheromone concentration of feature i at time t; is the pheromone concentration of feature i at time t+1; ρ(t) is the pheromone volatility rate at time t; m is the number of ants; is the weighted contribution of ant k to feature i at time t; the weighted contribution formula is: L k is the path length of ant k; Q is a constant representing the pheromone intensity, which is used to adjust the scale of pheromone update; d i is the importance score of feature i; D max is the highest importance score among all features; w k is the weighting factor of ant k, which can be adjusted according to the path quality of ant k; β is a parameter that controls the speed of change of the weighting factor;

[0128] The dynamic heuristic information update formula used to adjust the heuristic information according to the feature performance in the iterative process is: is the dynamic heuristic information of feature i in time; is the initial heuristic information of feature i; γ is a dynamic adjustment coefficient for controlling the degree of dynamic change of heuristic information; is the importance score of feature i at time t; f max is the highest importance score among all features; represents the importance score of feature i calculated using the random forest model M;

[0129] The adjustment formula of the adaptive cooling coefficient is: α(t) is the cooling coefficient of the tth iteration; α init is the preset initial cooling coefficient; β is the cooling coefficient decay rate, which determines the speed of the cooling coefficient decreasing with the increase of the iteration number; t is the current iteration number; MaxIter is the maximum iteration number.

[0130] In this embodiment, an optimization method integrating ant colony algorithm and simulated annealing algorithm is used to solve the selection problem of key electrical parameter features, so as to more accurately and efficiently identify the features highly related to fault diagnosis. Specifically, the probability formula defines the probability of ants selecting features, in which the pheromone concentration and heuristic information represent the attractiveness of features and the relevance to fault diagnosis respectively, and two preset parameters (α and β) are used to control the weights of the two factors in decision-making. The acceptance probability formula: in the simulated annealing algorithm, this formula defines the probability of accepting new solutions, in which the difference between the evaluation function values of the current solution and the candidate solution and the current temperature value are considered, which helps to balance the relationship between local optimum and global optimum in the search process. For the pheromone updating mechanism: first, the local updating formula is used to update the pheromone concentration after each iteration according to the quality of the path selected by ants, in which the importance score of features is introduced to reflect the contribution of features to fault diagnosis, and the pheromone evaporation rate is adjusted according to the path length and the current best solution length. In addition, the global updating formula is performed at the end of iteration, which adjusts the pheromone concentration based on the path length of ants and the importance score of features, and also considers the weighted factor of ants to further emphasize the features on excellent paths. And for the dynamic updating of heuristic information: the heuristic information is dynamically adjusted based on the importance score of features and its change over time, which helps the algorithm pay more attention to features showing better diagnostic performance in the iteration process. For the adaptive cooling coefficient: by using an adjustment formula to automatically change the cooling coefficient in the simulated annealing algorithm, the needs of global search and local optimization are better balanced, ensuring that the algorithm can extensively explore in the early stage and focus on better solutions in the later stage.

[0131] Therefore, the embodiment combines the ant colony algorithm and the simulated annealing algorithm to exert their respective advantages, improve the efficiency and accuracy of feature selection, and adaptively cope with different diagnostic scenarios by dynamically adjusting pheromone concentration, heuristic information and cooling coefficient, thereby improving the overall performance. The algorithm further enhances the recognition ability of relevant features by considering the importance score of features to adjust the pheromone concentration and heuristic information.

[0132] As an improvement of the above scheme, the feature fusion of the screened suspicious fault features is performed to obtain fused suspicious fault features, including:

[0133] The screened suspicious fault features are input into a feature fusion algorithm formula for feature fusion to obtain fused suspicious fault features; the feature fusion algorithm formula is:

[0134] F fusion is a fused feature vector; f i is an original feature vector of the i th suspicious fault feature; g i = σ(W g · f i +b g ) is a vector of the i th suspicious fault feature after nonlinear transformation; W g and b g are weights and biases of the nonlinear transformation layer; σ is a ReLU or Sigmoid activation function; score(g i ) = W s · σ(W g · g i +b g )+b s is a score of the i th suspicious fault feature, used for calculating an adaptive weight; W s and b s are weights and biases of the score calculation layer; exp is an exponential function, used for calculating an attention weight; j represents an index of a feature vector, is used for traversing the feature vector of each suspicious fault feature, calculating a sum of exponential function values of scores of each property after nonlinear transformation, thereby calculating a weight of the feature vector of each suspicious fault feature, and ensuring that the sum of the weights is 1; N represents a total number of the feature vectors of the suspicious fault features.

[0135] In this embodiment, the filtered suspected fault features are further processed using a feature fusion algorithm to obtain more accurate and effective fused suspected fault features. Specifically, the innovative feature fusion algorithm formula of this embodiment defines a feature fusion process that includes a nonlinear transformation layer and a score calculation layer. The nonlinear transformation layer performs a nonlinear transformation on the original feature vector of each suspected fault feature to capture the nonlinear relationships between features. The score calculation layer calculates the score of each suspected fault feature and then calculates an adaptive weight to reflect the contribution of each feature to the fusion result. The nonlinear transformation layer performs a nonlinear transformation on the original feature vector of the suspected fault feature using weight and bias parameters, using ReLU or Sigmoid activation functions to introduce nonlinear relationships and enhance the model's expressiveness. The score calculation layer calculates the score of each suspected fault feature, which is used to calculate the attention weight. The score is obtained through another nonlinear transformation layer, and the attention weight is calculated using an exponential function to ensure that the sum of the weights of all suspected fault features is 1. The adaptive weight calculation uses the exponential function and score calculation to assign an adaptive weight to each suspected fault feature. These weights reflect the relative importance of each suspected fault feature to the final fusion result. Therefore, this embodiment can more comprehensively consider the information of multiple suspected fault features through the feature fusion algorithm, thereby improving the accuracy of fault diagnosis.

[0136] As an improvement to the above solution, the trained deep learning model is used to analyze the fused suspected fault features to obtain the fault type of the intelligent weak current system, including:

[0137] The suspected fault features are input into the input layer of a deep learning model that is a hybrid model including multi-scale convolution, attention mechanism, GRU and adaptive learning rate. After the multi-scale extraction of local features by the multi-scale convolution layer of the deep learning model, feature dependency learning by the attention layer, long-term dependency capture by the GRU layer and classification by the fully connected layer, the output layer of the trained deep learning model predicts and outputs the fault type of the intelligent weak current system;

[0138] Deep learning models include:

[0139] An input layer for receiving preprocessed fusion feature data;

[0140] Multi-scale convolution layer for extracting multi-scale local features; the multi-scale convolution layer uses multiple filters where K k is the size of the kth filter, F is the number of filters; the convolution operation is expressed as: h (k) =f(W (k) *x+b (k) ), * represents convolution operation; b(k) is a bias vector; the convolution outputs h (k) are concatenated to obtain the final convolution output H;

[0141] is an attention layer for learning the interdependence between features; the attention weight of the attention layer is represented as: a = softmax(U·f(V·H+c)+d), U, V, c, d are weight and bias parameters; a is an attention weight vector; H is a convolution output;

[0142] is a GRU layer for processing sequence data and capturing long-term dependencies; the GRU layer uses a gated recurrent unit to process sequence data, represented as:

[0143] z t = σ(U z h t-1 +W z (H t ☉a t )+b z )

[0144] r t = σ(U r h t-1 +W r (H t ☉a t )+b r )

[0145]

[0146] where z t , r t and h t are the update gate, the reset gate and the hidden state, respectively; U z , U r , U h , W z , W r , W h are preset weights, b z , b r , b h are preset bias parameters; σ is a sigmoid activation function; tanh is a hyperbolic tangent function; H t is the output of the convolution layer; a t is the attention weight;

[0147] is a fully connected layer for a classification task; the fully connected layer is: z = f(Uh T +c), U is a weight matrix; c is a bias vector; h T is the GRU hidden state at the last time step;

[0148] an output layer for predicting the output fault type; the output layer uses a softmax function for classification: V is a preset weight matrix; d is a preset bias vector; is a preset prediction probability vector; z is a classification output of the full connection layer.

[0149] In this embodiment, the deep learning model is composed of multiple components, including an input layer, a multi-scale convolutional layer, an attention layer, a GRU layer, and a full connection layer, and finally a prediction of the fault type is made through an output layer. Among them, there are the following technical improvements: 1. Design of multi-scale convolutional layer: Extract multi-scale local features through filters of different sizes, enhancing the feature capturing ability of the model. 2. Application of attention mechanism: Highlight important features by learning attention weights, improving the model's attention to key information. 3. Use of GRU layer: Effectively process sequence data, capture long-term dependencies, and improve the model's time series processing capability. 4. Adaptive learning rate: Adjust the learning rate to speed up the training process and improve the model's convergence speed. Therefore, the deep learning model of this embodiment captures local features of different scales through the multi-scale convolutional layer, enhancing the model's ability to represent complex features. The attention mechanism can highlight important feature information, and the GRU layer can effectively process time series data, improving the model's generalization ability. In addition, the final classification task is performed through the full connection layer and the output layer, which can more accurately predict the type of fault occurrence. The fault types can include overload fault, short circuit fault, open circuit fault, temperature abnormality fault, and mechanical vibration fault, etc. The deep learning model can be trained using existing model training methods and training data such as fault type samples and fusion feature samples.

[0150] Referring to Figure 2 is a structural schematic diagram of an intelligent weak current system operation monitoring system provided by an embodiment of the application. The intelligent weak current system operation monitoring system comprises:

[0151] A receiving module 10 is configured to receive key electrical parameters of each target device in the intelligent weak current system collected by a data acquisition device in real time, to obtain a key electrical parameter set; the key electrical parameters include current, voltage, temperature, sound parameters, and vibration analysis parameters.

[0152] A feature extraction module 11 is configured to perform feature extraction on each parameter in the key electrical parameter set, to obtain an extracted key electrical parameter feature set; the key electrical parameter feature set includes current features, voltage features, temperature features, sound parameter features, and vibration analysis parameter features.

[0153] The feature screening module 12 is configured to screen suspicious fault features from the key electrical parameter feature set by combining an ant colony algorithm with a simulated annealing algorithm.

[0154] The feature fusion module 13 is configured to perform feature fusion on the screened suspicious fault features to obtain fused suspicious fault features.

[0155] The fault analysis module 14 is configured to analyze the fused suspicious fault features by using the trained deep learning model to obtain the fault occurrence type of the intelligent weak current system.

[0156] In the embodiment of the application, the data acquisition device is used to collect the key electrical parameters of each target device in the intelligent weak current system in real time to form a key electrical parameter set. Then, feature extraction is performed on these parameters to construct a key electrical parameter feature set. Next, the ant colony algorithm and the simulated annealing algorithm are combined to screen suspicious fault features from the feature set. Then, feature fusion is performed on the screened suspicious fault features to generate fused suspicious fault features. Finally, the fused suspicious fault features are analyzed by using the pre-trained deep learning model to accurately determine the fault type of the intelligent weak current system. In the embodiment of the application, the ant colony algorithm and the simulated annealing algorithm are combined to select features, and the deep learning model is used to diagnose the fused suspicious fault features, which significantly improves the accuracy and efficiency of fault detection.

[0157] As an improvement of the above-mentioned scheme, the feature screening module is specifically configured to:

[0158] perform ant colony algorithm parameter initialization and simulated annealing algorithm parameter initialization;

[0159] randomly select a part of the key electrical parameter features from the key electrical parameter feature set as an initial solution;

[0160] each ant selects a key electrical parameter feature according to the current pheromone concentration τ ij and the heuristic information η ij and uses a probability formula p ij to select a key electrical parameter feature, construct a feature subset, and construct all the feature subsets to form a solution space of the current iteration; the higher the pheromone concentration, the greater the probability of selection of the related key electrical parameter feature, and the more relevant the key electrical parameter feature to fault diagnosis;

[0161] randomly select a feature subset from the current solution space as a current solution;

[0162] add or delete a key electrical parameter feature in the current solution to generate a new feature subset, and use the new feature subset as a new solution;

[0163] The evaluation function value E' of the new solution is evaluated and compared with the current solution E; if E' < E, the new solution is accepted; if E' ≥ E, it is determined according to a preset acceptance probability formula whether to accept the new solution; the evaluation function is used to evaluate the effectiveness of different characteristic subsets, and the higher the evaluation function value, the more likely the characteristic subset contains suspicious fault characteristics;

[0164] After each iteration, the ants locally update the pheromone according to the path quality of the path selected by them using an improved local update formula;

[0165] At the end of each iteration, the pheromone is updated according to the selection results of all ants and through an improved global update formula, and the high-quality path is strengthened;

[0166] The heuristic information is adjusted according to the characteristic performance in the iteration process, guiding the ants to preferentially select high-quality key electrical parameter characteristics; the higher the heuristic information, the higher the probability of the related key electrical parameter characteristic being selected as a suspicious fault characteristic;

[0167] The adaptive cooling coefficient in the simulated annealing algorithm is adjusted to better balance the relationship between global search and local optimization;

[0168] When the preset iteration stopping condition is met, the algorithm ends and outputs the final characteristic combination, so that a suspicious fault characteristic is obtained from the key electrical parameter characteristic set.

[0169] As an improvement of the above-mentioned scheme, the probability formula p ij is: p i is the probability of the ant k selecting the characteristic i; τ i is the pheromone concentration of the characteristic i; η i is the heuristic information of the characteristic i; and α and β are preset algorithm parameters, respectively controlling the influence degree of the pheromone and the heuristic information;

[0170] The acceptance probability formula is: P(e, e', T is the probability of accepting the new solution; E is the evaluation function value of the current solution; E' is the evaluation function value of the candidate solution; and T is the current temperature;

[0171] After each iteration, the ants locally update the pheromone according to the path quality of the path selected by them using an improved local update formula: τ is the pheromone concentration of the characteristic i at time t; τ is the pheromone concentration of the characteristic i at time t+1; ρ(t) is the pheromone evaporation rate at time t; Q is a constant representing pheromone intensity, used to adjust the scale of pheromone update; L best is the length of the current best solution; d i is the importance score of feature i; D max is the highest importance score among all features; ρ0is the initial evaporation rate; p max is the maximum evaporation rate; L best is the path length of the current best solution; L avg is the path length of the current average solution; γ is a parameter controlling the speed of evaporation rate change;

[0172] The improved global update formula is: is the pheromone concentration of feature i at time t; is the pheromone concentration of feature i at time t+1; ρ(t) is the pheromone evaporation rate at time t; m is the number of ants; is the weighted contribution of ant k to feature i at time t; the weighted contribution formula is: L k is the path length of ant k; Q is a constant representing pheromone intensity, used to adjust the scale of pheromone update; d i is the importance score of feature i; D max is the highest importance score among all features; w k is the weighting factor of ant k, which can be adjusted according to the path quality of ant k; β is a parameter controlling the speed of weighting factor change;

[0173] The dynamic heuristic information update formula used to adjust the heuristic information according to the feature performance in the iteration process is: is the dynamic heuristic information of feature i at time t; is the initial heuristic information of feature i; γ is a dynamic adjustment coefficient used to control the degree of dynamic change of heuristic information; is the importance score of feature i at time t; f max is the highest importance score among all features; represents the importance score of feature i calculated using the random forest model M;

[0174] The adjustment formula of the adaptive cooling coefficient is: α(t) is the cooling coefficient of the tth iteration; α init is the preset initial cooling coefficient; β is the cooling coefficient decay rate, which determines the speed of cooling coefficient decrease with the increase of iteration number; t is the current iteration number; MaxIter is the maximum iteration number.

[0175] As an improvement of the above scheme, the feature fusion module is specifically used for:

[0176] The screened suspicious fault features are input into the feature fusion algorithm formula for feature fusion to obtain fused suspicious fault features; the feature fusion algorithm formula is:

[0177] F fusion is a fused feature vector; f i is an original feature vector of the i-th suspicious fault feature; g i = σ(W g · f i +b g ) is a vector after nonlinear transformation of the i-th suspicious fault feature; W g and b g are weights and biases of the nonlinear transformation layer; σ is a ReLU or Sigmoid activation function; score(g i ) = W s · σ(W g · g i +b g )+b s is a score of the i-th suspicious fault feature, used for calculating an adaptive weight; W s and b s are weights and biases of the score calculation layer; exp is an exponential function, used for calculating an attention weight; j represents an index of a feature vector, for traversing a feature vector of each suspicious fault feature, calculating a sum of exponential function values of scores after nonlinear transformation of each property, thereby calculating a weight of the feature vector of each suspicious fault feature, and ensuring that the sum of the weights is 1; N represents a total number of the feature vectors of the suspicious fault features.

[0178] As an improvement of the above scheme, the fault analysis module is specifically used for:

[0179] The suspicious fault features are input into an input layer of a deep learning model of a hybrid model containing multi-scale convolution, attention mechanism, GRU and adaptive learning rate, and after sequentially passing through local feature multi-scale extraction of a multi-scale convolution layer of the deep learning model, feature dependency relationship learning of an attention layer, long-term dependency relationship capture of a GRU layer and classification of a full connection layer, a fault occurrence type of the intelligent weak current system is predicted by an output layer of the trained deep learning model.

[0180] The deep learning model comprises:

[0181] an input layer for receiving preprocessed fused feature data;

[0182] a multi-scale convolutional layer for extracting multi-scale local features; the multi-scale convolutional layer uses multiple filters where K k is the size of the kth filter, F is the number of filters; the convolution operation is represented as: h (k) = f(W (k) *x+b (k) ), * represents the convolution operation; b (k) is a bias vector; the convolution outputs of different scales h (k) are spliced to obtain the final convolution output H;

[0183] an attention layer for learning the mutual dependence between features; the attention weight of the attention layer is represented as: a = softmax(U·f(V·H+c)+d), U, V, c, d are weight and bias parameters; a is an attention weight vector; H is the convolution output;

[0184] a GRU layer for processing sequence data and capturing long-term dependencies; the GRU layer uses a gated recurrent unit to process sequence data, represented as:

[0185] z t = σ(U z h t-1 +W z (H t ☉a t )+b z )

[0186] r t = σ(U r h t-1 +W r (H t ☉a t )+b r )

[0187]

[0188] where z t , r t , and h t are the update gate, the reset gate, and the hidden state, respectively; U z , U r , U h , W z , W r , W h are preset weights, b z , b r , b h are preset bias parameters; σ is a sigmoid activation function; tanh is a hyperbolic tangent function; H tis the output of the convolutional layer; a t is the attention weight;

[0189] is the output of the fully connected layer for the classification task; the fully connected layer is: z = f(Uh T +c), U is a weight matrix; c is a bias vector; h T is the GRU hidden state of the last time step;

[0190] is the output layer for predicting the output fault type; the output layer uses a softmax function for classification: V is a preset weight matrix; d is a preset bias vector; is a preset prediction probability vector; z is the classification output of the fully connected layer.

[0191] It should be noted that the system embodiments described above are only illustrative, and the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. In addition, the connection relationship between the modules in the device embodiment provided by the present application indicates that there is a communication connection between them, which can be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement it without creative labor.

[0192] The above is the preferred embodiment of the present application. It should be noted that those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements are also considered within the scope of protection of the present application.

Claims

1. A method for monitoring the operation of an intelligent weak current system, applied to a cloud server, characterized in that: It includes: Receiving the key electrical parameters of each target device in the intelligent weak current system collected by the data acquisition device in real time, and obtaining a set of key electrical parameters; The key electrical parameters include: current, voltage, temperature, sound parameters, and vibration analysis parameters; Performing feature extraction on each parameter of the set of key electrical parameters to obtain an extracted set of key electrical parameter features; the set of key electrical parameter features includes current features, voltage features, temperature features, sound parameter features, and vibration analysis parameter features; Screening for suspicious fault features from the set of key electrical parameter features by combining the ant colony algorithm and the simulated annealing algorithm; Performing feature fusion on the screened suspicious fault features to obtain fused suspicious fault features; Analyzing the fused suspicious fault features through a trained deep learning model to obtain the type of fault occurrence in the intelligent weak current system.

2. The operation monitoring method of the intelligent weak current system according to claim 1, characterized in that: The screening for suspicious fault features from the set of key electrical parameter features by combining the ant colony algorithm and the simulated annealing algorithm includes: Initializing the parameters of the ant colony algorithm and the parameters of the simulated annealing algorithm; Randomly selecting a part of the key electrical parameter features from the set of key electrical parameter features as the initial solution; Each ant has a current pheromone concentration τ ij and heuristic information η ij And use the probability formula p ij Select key electrical parameter features, construct feature subsets, and use all constructed feature subsets to form the solution space for the current iteration. The higher the pheromone concentration, the greater the probability that the relevant key electrical parameter feature will be selected, and the more relevant this key electrical parameter feature is to fault diagnosis. Randomly selecting a feature subset from the current solution space as the current solution; Generating a new feature subset by randomly adding or deleting a key electrical parameter feature in the current solution, and taking this new feature subset as the new solution; Evaluating the evaluation function value E′ of the new solution and comparing it with the current solution E; if E′ < E, accept the new solution; if E′ ≥ E, decide whether to accept the new solution according to a preset acceptance probability formula; the evaluation function is used to evaluate the effectiveness of different feature subsets, and the higher the evaluation function value, the more likely the feature subset contains suspicious fault features; After each iteration, the ants locally update the pheromone according to the path quality of the paths they choose using an improved local update formula; At the end of each iteration, update the pheromone according to the selection results of all ants and through an improved global update formula to strengthen the high-quality paths; Adjust the heuristic information according to the feature performance during the iteration process to guide the ants to be more inclined to select high-quality key electrical parameter features; the higher the heuristic information, the relatively higher the probability that the relevant key electrical parameter features are selected as suspicious fault features; Adjust the adaptive cooling coefficient in the simulated annealing algorithm to better balance the relationship between global search and local optimization; When the preset iteration stop condition is met, the algorithm ends and outputs the final feature combination, so as to obtain suspicious fault features from the set of key electrical parameter features.

3. The operation monitoring method of the intelligent weak current system according to claim 2, wherein The probability formula p ij for: p i is the probability that ant k selects feature i; τ i is the pheromone concentration of feature i; η i is the heuristic information of feature i; α and β are preset algorithm parameters that control the influence of pheromone and heuristic information respectively; The acceptance probability formula is: P(e, e′, T) is the probability of accepting a new solution; E is the evaluation function value of the current solution; E′ is the evaluation function value of the candidate solution; T is the current temperature; After each iteration, the ant locally updates the pheromone according to the path quality of the path it chooses. The improved local update formula is: is the pheromone concentration of feature i at time t; is the pheromone concentration of feature i at time t+1; ρ(t) is the pheromone volatilization rate at time t; is the contribution of ant k to feature i at time t; the contribution formula is: Q is a constant representing the pheromone intensity, which is used to adjust the scale of pheromone update; L best is the length of the current best solution; d i is the importance score of feature i; D max is the highest importance score among all features; ρ0 is the initial volatility; ρ max is the maximum volatility; L best is the path length of the current best solution; L avg is the path length of the current average solution; γ is the parameter that controls the rate of change of the volatility; The improved global update formula is: is the pheromone concentration of feature i at time t; is the pheromone concentration of feature i at time t+1; ρ(t) is the pheromone volatility rate at time t; m is the number of ants; is the weighted contribution of ant k to feature i at time t; The weighted contribution formula is: L k is the path length of ant k; Q is a constant representing the pheromone intensity, which is used to adjust the scale of pheromone update; d i is the importance score of feature i; D max is the highest importance score among all features; w k is the weighting factor of ant k, which can be adjusted according to the path quality of ant k; β is a parameter that controls the speed of change of the weighting factor; The dynamic heuristic information update formula used to adjust the heuristic information according to the feature performance in the iterative process is: is the dynamic heuristic information of feature i in time; is the initial heuristic information of feature i; γ is the dynamic adjustment coefficient, which is used to control the dynamic change degree of heuristic information; is the importance score of feature i at time t; f max is the highest importance score among all features; represents the importance score of feature i calculated using the random forest model M; The adjustment formula of the adaptive cooling coefficient is: α(t) is the cooling coefficient of the tth iteration; α init is the preset initial cooling coefficient; β is the cooling coefficient decay rate, which determines the speed at which the cooling coefficient decreases with the increase of the number of iterations; t is the current number of iterations; MaxIter is the maximum number of iterations.

4. The operation monitoring method of the intelligent weak current system according to claim 1, characterized in that: The performing feature fusion on the screened suspicious fault features to obtain fused suspicious fault features includes: Inputting the screened suspicious fault features into the feature fusion algorithm formula for feature fusion to obtain fused suspicious fault features; the feature fusion algorithm formula is: F fusion is the fused feature vector; f i is the original feature vector of the i-th suspected fault feature; g i =σ(W g ·f i +b g ) is the vector of the i-th suspected fault feature after nonlinear transformation; W g and b g is the weight and bias of the nonlinear transformation layer; σ is the ReLU or Sigmoid activation function; score(g i )=W s ·σ(W g ·g i +b g )+b s is the score of the i-th suspected fault feature, which is used to calculate the adaptive weight; W s and b s are the weights and biases of the score calculation layer; exp is the exponential function used to calculate the attention weight; j represents the index of the feature vector, It is used to traverse the eigenvector of each suspected fault feature, calculate the sum of the exponential function values ​​of the scores of each property after nonlinear transformation, and thus calculate the weight of the eigenvector of each suspected fault feature to ensure that the sum of the weights is 1; N represents the total number of eigenvectors of suspected fault features.

5. The operation monitoring method of the intelligent weak current system according to claim 1, characterized in that: The analyzing the fused suspicious fault features through a trained deep learning model to obtain the type of fault occurrence in the intelligent weak current system includes: The suspected fault features are input into the input layer of a deep learning model that is a hybrid model including multi-scale convolution, attention mechanism, GRU and adaptive learning rate. After the multi-scale extraction of local features by the multi-scale convolution layer of the deep learning model, feature dependency learning by the attention layer, long-term dependency capture by the GRU layer and classification by the fully connected layer, the output layer of the trained deep learning model predicts and outputs the fault type of the intelligent weak current system; Deep learning models include: An input layer for receiving preprocessed fusion feature data; Multi-scale convolution layer for extracting multi-scale local features; the multi-scale convolution layer uses multiple filters where K k is the size of the kth filter, F is the number of filters; the convolution operation is expressed as: h (k) =f(W (k) *x+b (k) ), * represents convolution operation; b (k) Is the bias vector; convolution output h of different scales (k) Splicing is performed to obtain the final convolution output H; An attention layer is used to learn the interdependencies between features. The attention weight of the attention layer is expressed as: a = softmax(U f(V H + c) + d), where U, V, c, and d are weight and bias parameters; a is the attention weight vector; and H is the convolution output. GRU layer for processing sequence data and capturing long-term dependencies; the GRU layer uses a gated recurrent unit to process sequence data, expressed as: where z t ,r t and h t They are update gate, reset gate and hidden state respectively; U z , U r , U h , W z , W r , W h is the preset weight, b z ,b r ,b h is the preset bias parameter; σ is the sigmoid activation function; tanh is the hyperbolic tangent function; H t is the output of the convolutional layer; a t is the attention weight; The fully connected layer for classification tasks is: z = f(Uh T +c), U is the weight matrix; c is the bias vector; h T is the GRU hidden state at the last time step; The output layer is used to predict the output fault type; the output layer uses the softmax function for classification: V is the preset weight matrix; d is the preset bias vector; is the preset prediction probability vector; z is the classification output of the fully connected layer.

6. An operation monitoring system for an intelligent weak current system, characterized in that: include: A receiving module is used to receive the key electrical parameters of each target device in the intelligent weak current system collected in real time by the data acquisition device to obtain a key electrical parameter set; The key electrical parameters include: current, voltage, temperature, sound parameters and vibration analysis parameters; a feature extraction module, configured to extract features from each parameter of the key electrical parameter set to obtain an extracted key electrical parameter feature set; the key electrical parameter feature set includes current features, voltage features, temperature features, sound parameter features, and vibration analysis parameter features; A feature screening module, configured to screen suspected fault features from the key electrical parameter feature set by combining an ant colony algorithm with a simulated annealing algorithm; The feature fusion module is used to fuse the screened suspicious fault features to obtain the fused suspicious fault features; The fault analysis module is used to analyze the fused suspicious fault features through the trained deep learning model to obtain the fault type of the intelligent weak current system.

7. The operation monitoring system of the intelligent weak current system according to claim 6, characterized in that: The feature screening module is specifically used for: Initialize the parameters of the ant colony algorithm and the simulated annealing algorithm; Randomly selecting a portion of key electrical parameter features from the key electrical parameter feature set as an initial solution; Each ant has a current pheromone concentration τ ij and heuristic information η ij And use the probability formula p ij Select key electrical parameter features, construct feature subsets, and use all constructed feature subsets to form the solution space for the current iteration. The higher the pheromone concentration, the greater the probability that the relevant key electrical parameter feature will be selected, and the more relevant this key electrical parameter feature is to fault diagnosis. Randomly select a feature subset from the current solution space as the current solution; Generating a new feature subset by randomly adding or deleting a key electrical parameter feature in the current solution, and using the new feature subset as a new solution; Evaluate the evaluation function value E′ of the new solution and compare it with the current solution E. If E′ < E, accept the new solution. If E′ ≥ E, decide whether to accept the new solution based on the preset acceptance probability formula. The evaluation function is used to evaluate the effectiveness of different feature subsets. The higher the evaluation function value, the more likely the feature subset contains suspicious fault features. After each iteration, the ants locally update the pheromone using the improved local update formula according to the path quality of their chosen path; At the end of each iteration, the pheromone is updated according to the selection results of all ants through the improved global update formula to strengthen the high-quality path; Heuristic information is adjusted based on the feature performance during the iteration process to guide ants to prefer high-quality key electrical parameter features. The higher the heuristic information, the higher the probability that the relevant key electrical parameter feature will be selected as a suspected fault feature. Adjusted the adaptive cooling coefficient in the simulated annealing algorithm to better balance the relationship between global search and local optimization; When a preset iteration stop condition is met, the algorithm ends and outputs a final feature combination, thereby obtaining a suspected fault feature from the key electrical parameter feature set.

8. The operation monitoring system of the intelligent weak current system according to claim 7, characterized in that: The probability formula p ij for: p i is the probability that ant k selects feature i; τ i is the pheromone concentration of feature i; η i is the heuristic information of feature i; α and β are preset algorithm parameters that control the influence of pheromone and heuristic information respectively; The acceptance probability formula is: P(e, e′, T) is the probability of accepting a new solution; E is the evaluation function value of the current solution; E′ is the evaluation function value of the candidate solution; T is the current temperature; After each iteration, the ant locally updates the pheromone according to the path quality of the path it chooses. The improved local update formula is: is the pheromone concentration of feature i at time t; is the pheromone concentration of feature i at time t+1; ρ(t) is the pheromone volatilization rate at time t; is the contribution of ant k to feature i at time t; the contribution formula is: Q is a constant representing the pheromone intensity, which is used to adjust the scale of pheromone update; L best is the length of the current best solution; d i is the importance score of feature i; D max is the highest importance score among all features; ρ0 is the initial volatility; ρ max is the maximum volatility; L best is the path length of the current best solution; L avg is the path length of the current average solution; γ is the parameter that controls the rate of change of the volatility; The improved global update formula is: is the pheromone concentration of feature i at time t; is the pheromone concentration of feature i at time t+1; ρ(t) is the pheromone volatility rate at time t; m is the number of ants; is the weighted contribution of ant k to feature i at time t; the weighted contribution formula is: L k is the path length of ant k; Q is a constant representing the pheromone intensity, which is used to adjust the scale of pheromone update; d i is the importance score of feature i; D max is the highest importance score among all features; w k is the weighting factor of ant k, which can be adjusted according to the path quality of ant k; β is a parameter that controls the speed of change of the weighting factor; The dynamic heuristic information update formula used to adjust the heuristic information according to the feature performance in the iterative process is: is the dynamic heuristic information of feature i in time; is the initial heuristic information of feature i; γ is the dynamic adjustment coefficient, which is used to control the dynamic change degree of heuristic information; is the importance score of feature i at time t; f max is the highest importance score among all features; represents the importance score of feature i calculated using the random forest model M; The adjustment formula of the adaptive cooling coefficient is: α(t) is the cooling coefficient of the tth iteration; α init is the preset initial cooling coefficient; β is the cooling coefficient decay rate, which determines the speed at which the cooling coefficient decreases with the increase of the number of iterations; t is the current number of iterations; MaxIter is the maximum number of iterations.

9. The operation monitoring system of the intelligent weak current system according to claim 6, characterized in that: The feature fusion module is specifically used for: Input the screened suspicious fault features into the feature fusion algorithm formula to perform feature fusion and obtain the fused suspicious fault features; feature The fusion algorithm formula is: F fusion is the fused feature vector; f i is the original feature vector of the i-th suspected fault feature; g i =σ(W g ·f i +b g ) is the vector of the i-th suspected fault feature after nonlinear transformation; W g and b g is the weight and bias of the nonlinear transformation layer; σ is the ReLU or Sigmoid activation function; score(g i )=W s ·σ(W g ·g i +b g )+b s is the score of the i-th suspected fault feature, which is used to calculate the adaptive weight; W s and b s are the weights and biases of the score calculation layer; exp is the exponential function used to calculate the attention weight; j represents the index of the feature vector, It is used to traverse the eigenvector of each suspected fault feature, calculate the sum of the exponential function values ​​of the scores of each property after nonlinear transformation, and thus calculate the weight of the eigenvector of each suspected fault feature to ensure that the sum of the weights is 1; N represents the total number of eigenvectors of suspected fault features.

10. The operation monitoring system of the intelligent weak current system according to claim 6, characterized in that: The fault analysis module is specifically used for: The suspected fault features are input into the input layer of a deep learning model that is a hybrid model including multi-scale convolution, attention mechanism, GRU and adaptive learning rate. After the multi-scale extraction of local features by the multi-scale convolution layer of the deep learning model, feature dependency learning by the attention layer, long-term dependency capture by the GRU layer and classification by the fully connected layer, the output layer of the trained deep learning model predicts and outputs the fault type of the intelligent weak current system; Deep learning models include: An input layer for receiving preprocessed fusion feature data; Multi-scale convolution layer for extracting multi-scale local features; the multi-scale convolution layer uses multiple filters where K k is the size of the kth filter, F is the number of filters; the convolution operation is expressed as: h (k) =f(W (k) *x+b (k) ), * represents convolution operation; b (k) Is the bias vector; convolution output h of different scales (k) Splicing is performed to obtain the final convolution output H; An attention layer is used to learn the interdependencies between features. The attention weight of the attention layer is expressed as: a = softmax(U f(V H + c) + d), where U, V, c, and d are weight and bias parameters; a is the attention weight vector; and H is the convolution output. GRU layer for processing sequence data and capturing long-term dependencies; the GRU layer uses a gated recurrent unit to process sequence data, expressed as: where z t , r t and h t They are update gate, reset gate and hidden state respectively; U z ,U r , U h , W z , W r , W h is the preset weight, b z , b r , b h is the preset bias parameter; σ is the sigmoid activation function; tanh is the hyperbolic tangent function; H t is the output of the convolutional layer; a t is the attention weight; The fully connected layer for classification tasks is: z = f(Uh T +c), U is the weight matrix; c is the bias vector; h T is the GRU hidden state at the last time step; The output layer is used to predict the output fault type; the output layer uses the softmax function for classification: V is the preset weight matrix; d is the preset bias vector; is the preset prediction probability vector; z is the classification output of the fully connected layer.

Images