Malfunction early-warning method for production logistics delivery equipment

Abstract

Disclosed is a malfunction early-warning method for production logistics delivery equipment. After a sensor obtains past signal data, performing feature extraction and dimensionality reduction so as to obtain a feature vector; using a growing neural gas (GNG) algorithm to divide normal state data into different operation situations so as to obtain several cluster centers, and calculating the Euclidean distance between the feature vector and the cluster centers obtained from current operation data, so as to obtain a similarity trend; constructing a past memory matrix, using an improved particle swarm algorithm to optimize an LS-SVM regression model parameter, and calculating the residual value of the current state. Finally, combining the residual value and the similarity trend to obtain a risk coefficient, assessing the equipment state, and issuing an early warning for an equipment malfunction.

Claims

1. A malfunction early warning method of production logistics delivery equipment, comprising the following steps: arranging a plurality of sensors in a production line, and collecting data including vibration acceleration signals of bearings and a speed reducer, belt displacement, of main parts of the equipment used in production are collected by a plurality of sensors; and detecting and predicting vibration acceleration for bearings and gearboxes, detecting and predicting current and voltage signals for servo motors, and detecting and predicting belt displacement signals, detecting and predicting temperature and humidity signals of major components, steps configured to be executed by one or more processors, comprising step 1, calculating a feature vector of a historical normal operation state based on signal data obtained by the plurality of sensors, dividing normal state data into a plurality of work conditions to obtain a plurality of clustering centers, and calculating a Euclidean distance from a current state to the clustering centers so as to obtain a similarity trend; step 2, building a historical memory matrix, optimizing parameters of an LS-SVM regression model, and calculating a residual of the current state and the regression model; and step 3, obtaining a risk coefficient by combining the similarity trend and the residual, evaluating an equipment operation state, and making timely early warning on faults; wherein the step 1 comprises the following specific processes: step 1.1, performing initialization: creating two nodes with weight vectors, and a zero value of a local error; step 1.2, inputting a vector into a neural network x, and finding two nerve cells s and t in positions closest to the x, i.e., the nodes with weight vectors w.sub.s and w.sub.t, wherein ∥w.sub.s−x∥.sup.2 is a node with a smallest distance value in all nodes, and ∥w.sub.t−x∥.sup.2 is a node with a second-smallest distance value in all nodes; step 1.3, updating a local error of a winner nerve cell s, and adding the local error of the winner nerve cell s into a squared distance of the vector w.sub.S and the x:
E.sub.S←E.sub.S+∥w.sub.S−x∥.sup.2  (1); step 1.4, translating the winner nerve cell s and all topological neighbors thereof, wherein a direction is an input vector x, and distances equal to partial ∈.sub.w and whole ∈.sub.n:
w.sub.s←w.sub.s+∈.sub.w.Math.(w.sub.s−x)  (2), and
w.sub.n←w.sub.n+∈.sub.n.Math.(w.sub.n−x)  (3); step 1.5, by using 1 as a step amplitude, increasing ages of all connections from the winner nerve cell s, and removing the connections with the ages being elder than age.sub.max; and if a result in the nerves cells does not have more divergence margins, also removing the nerve cells; step 1.6, if the number of current iterations is a multiple of λ, and does not reach a limit dimension of a network, inserting a new nerve cell r as follows; step 1.7, reducing all errors of a nerve cell j by using a fraction β:
E.sub.j←E.sub.j−E.sub.j.Math.β  (4); and step 1.8, if a stop condition is not met, continuing the step 2.

2. The malfunction early warning method of production logistics delivery equipment according to claim 1, wherein in the step 2, the improved particle swarm algorithm is used to optimize a kernel function σ and a penalty coefficient c in the LS-SVM regression model.

3. The malfunction early warning method of production logistics delivery equipment according to claim 2, wherein the step 2 comprises the following specific processes: step 2.1.1, building the LS-SVM regression model: introducing a Lagrangian function for solving it, and selecting a radial basis function K(x,x.sub.i)=exp(−∥x−x.sub.i∥.sub.2/2σ.sup.2), wherein σ is a kernel width; and obtaining the LS-SVM regression model: $\begin{array}{cc}f\left(x\right)=\underset{i=1}{\overset{l}{.Math.}}{\alpha }_{i}K\left(x,x_i\right)+b;& \left(5\right)\end{array}$ step 2.1.2, checking whether a historical optimum adaptive degree P.sub.b meets a constraint condition or whether the number of iterations reaches the maximum, if the constraint condition is still not met, and the number of iterations is not the maximum, performing a step 2.1.3, and otherwise, mapping a result into the kernel function σ and the penalty coefficient c of the LS-SVM model; and step 2.1.3, regulating speed and positions of particles, and regulating an inertia weight.

4. The malfunction early warning method of production logistics delivery equipment according to claim 3, wherein in the step 2.1.3, a self-adaptative regulation inertia weight method is used to regulate the inertia weight: $\begin{array}{cc}w=\{\begin{array}{cc}{w}_{\min }-\frac{\left(w_{\max }-w_{\min }\right)\left(f-f_{\min }\right)}{f_{\mathrm{avg}}-f_{\min }},& f<f_{avg}\\ w_{\max },& f\ge f_{\mathrm{avg}}\end{array}& \left(10\right)\end{array}$ wherein in the formula, w.sub.min is a minimum value of w, w.sub.max is a maximum value of w, f is an adaptive degree of a current particle, f.sub.avg is an average adaptive value of all particles, and f.sub.min is a minimum adaptive value of all particles.

5. The malfunction early warning method of production logistics delivery equipment according to claim 1, wherein the step 3 comprises the following specific processes: step 3.1, calculating a residual r.sub.i of the current state; step 3.2, calculating a similarity trend t.sub.i of the current state; and step 3.3, calculating a risk coefficient d.sub.i.

6. The malfunction early warning method of production logistics delivery equipment according to claim 5, wherein a specific process for calculating the residual r.sub.i of the current state in the step 3.1 is as follows:
r.sub.i=y.sub.i−f(x.sub.i)  (11) wherein in the formula, y.sub.i is a true value in a sample set, and f(x.sub.i) is a predicated value of the LS-SVM regression model after optimization by the improved particle swarm algorithm.

7. The malfunction early warning method of production logistics delivery equipment according to claim 5, wherein a specific process for calculating the similarity trend t.sub.i of the current state in the step 3.2 is as follows: $\begin{array}{cc}{t}_i=\frac{1}{n}\sqrt{\underset{j=1}{\overset{n}{.Math.}}{\left(x_i-X_j\right)}^2}& \left(12\right)\end{array}$ wherein in the formula, x.sub.i is a coordinate of the current state, and X.sub.j is a coordinate of the j-th clustering center.

8. The malfunction early warning method of production logistics delivery equipment according to claim 5, wherein a specific process for calculating the risk coefficient d.sub.i in the step 3.3 is as follows:
d.sub.i=ar.sub.i+bt.sub.i  (13) wherein in the formula, a and b are weight factors, and are initialized to 0.5 and 0.5 according to historical data.

9. The malfunction early warning method of production logistics delivery equipment according to claim 1, wherein the step 1.6 comprises the following specific processes: step 1.6.1, determining a nerve cell u with a greatest local error; step 1.6.2, determining a nerve cell v, with a greatest error, of the u in neighbors; step 1.6.3, creating a “middle” node r in the middle between the u and the v: $\begin{array}{cc}{w}_r=\frac{w_u+w_v}{2};& \left(6\right)\end{array}$ step 1.6.4, replacing an edge between the u and the v with an edge between the u and the r and an edge between the v and the r; and step 1.6.5, reducing the errors of the nerve cells u and v, and setting an error value of the nerve cell r:
E.sub.u←E.sub.u.Math.a  (7),
E.sub.v←E.sub.v.Math.a  (8), and
E.sub.r←E.sub.u  (9).

10. The malfunction early warning method of production logistics delivery equipment according to claim 1, wherein the step of translating the winner nerve cell s and all topological neighbors thereof refers to all nerve cells having connections with the winner nerve cell s.

11. The malfunction early warning method of production logistics delivery equipment according to claim 1, wherein in the step 1.5, if the two optimum nerve cells s and t are connected, the age of the connection is set to be zero, and otherwise, a connection is created between the two optimum nerve cells.

12. The malfunction early warning method of production logistics delivery equipment according to claim 1, wherein in the step 1, a growing neural gas (GNG) algorithm is used to calculate the feature vector of the historical normal operation state.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 is a flow diagram of an embodiment of the present invention.

DETAILED DESCRIPTION OF THE INVENTION

(2) The technical scheme of the present invention is illustrated in detail in conjunction with the drawings.

(3) Referring to FIG. 1, in order to realize the uninterrupted collection and processing of signal data, a large number of smart sensors are arranged in the production line. All sensors need to adopt active power supply and be installed inside the metal box to enhance the electromagnetic shielding ability. After the data is collected, it is continuously transmitted to the designated port of the data center server through the optical fiber ring network of the automobile assembly workshop for 24 hours.

(4) The present embodiment illustrates a malfunction early warning method of delivery equipment for automobile assembly of the present invention by taking the equipment used in automobile assembly line production as an example. As shown in FIG. 1, the method includes the following steps: Step 1, data collection: state data, including vibration acceleration signals of two bearings and a speed reducer, belt displacement and the like, of main parts of the equipment used in production are collected by a sensor. Step 2, feature parameter extraction: feature extraction is performed on different data by using different feature extraction technologies. Step 3, data dimension reduction: an effective value and a peak value of vibration removal are averaged. Then, a feature vector with a dimension being 7 is synthesized. The steps are repeated to obtain a plurality of feature vectors with the dimension being 7; Step 4, building of a growing neural gas (GNG) neural network model: for data with a historical normal operation state, a GNG algorithm is used to divide normal state data into a plurality of work conditions to obtain a plurality of clustering centers, and a Euclidean distance from a current state to the clustering centers is calculated so as to obtain a similarity trend. Step 5, building of an LS-SVM regression model: a historical memory matrix is built. Parameters of the LS-SVM regression model are optimized by an improved particle swarm algorithm. A residual of the current state and the regression model is calculated. Step 6, risk coefficient calculation: a risk coefficient is obtained by combining the similarity trend and the residual. An equipment operation state is evaluated. Timely early warning is made on faults.

(5) A malfunction early warning method of production logistics delivery equipment includes the following steps: Step 1, a feature vector of a historical normal operation state is calculated. Normal state data is divided into a plurality of work conditions to obtain a plurality of clustering centers. A Euclidean distance from a current state to the clustering centers is calculated so as to obtain a similarity trend. Step 2, a historical memory matrix is built. Parameters of an LS-SVM regression model are optimized by an improved particle swarm algorithm. A residual of the current state and the regression model is calculated. Step 3, a risk coefficient is obtained by combining with the similarity trend and the residual. An equipment operation state is evaluated. Timely early warning is made on faults.

(6) Further, the step 1 includes the following specific processes: Step 1.1, initialization is performed. Specifically, two nodes with weight vectors, and a zero value of a local error are created. Step 1.2, a vector is input into a neural network x. Two nerve cells s and t in positions closest to the x, i.e., the nodes with weight vectors w.sub.s and w.sub.t, are found. ∥w.sub.s−x∥.sup.2 is a node with a smallest distance value in all nodes, and ∥w.sub.t−x∥.sup.2 is a node with a second-smallest distance value in all nodes. Step 1.3, a local error of the winner nerve cell s is updated, and the local error of the winner nerve cell s is added into a squared distance of the vector w.sub.s and the x:
E.sub.s←E.sub.s+∥w.sub.s−x∥.sup.2  (1). Step 1.4, the winner nerve cell s and all topological neighbors thereof are translated. An direction is an input vector x, and distances equal to partial ∈.sub.w and whole ∈.sub.n:
w.sub.s←w.sub.s+∈.sub.s.Math.(w.sub.s−x)  (2), and
w.sub.n←w.sub.n+∈.sub.n.Math.(w.sub.n−x)  (3); Step 1.5, by using 1 as a step amplitude, ages of all connections from the winner nerve cell s are increased, and the connections with the ages being elder than age.sub.max are removed. If a result in the nerves cells does not have more divergence margins, the nerve cells are also removed. Step 1.6, if the number of current iterations is a multiple of λ, and does not reach a limit dimension of a network, a new nerve cell r is inserted as follows. Step 1.7, all errors of a nerve cell j are reduced by using a fraction β:
E.sub.j←E.sub.j−E.sub.j.Math.β  (4). Step 1.8, if a stop condition is not met, the step 2 is continued.

(7) Further, in the step 2, the improved particle swarm algorithm is used to optimize a kernel function σ and a penalty coefficient c in the LS-SVM regression model.

(8) Further, the step 3 includes the following specific processes: Step 3.1, a residual r.sub.i of the current state is calculated. Step 3.2, a similarity trend t.sub.i of the current state is calculated. Step 3.3, a risk coefficient d.sub.i is calculated.

(9) Further, the step 1.6 includes the following specific processes: Step 1.6.1, a nerve cell u with a greatest local error is determined. Step 1.6.2, a nerve cell v, with a greatest error, of the u in neighbors is determined. Step 1.6.3, a “middle” node r is created in the middle between the u and the v:

(10) $\begin{array}{cc}{w}_r=\frac{w_u+w_v}{2}.& \left(6\right)\end{array}$ Step 1.6.4, an edge between the u and the v is replaced with an edge between the u and the r and an edge between the v and the r. Step 1.6.5, the errors of the nerve cells u and v are reduced, and an error value of the nerve cell r is set:
E.sub.u←E.sub.u.Math.a  (7),
E.sub.v←E.sub.v.Math.a  (8), and
E.sub.r<E.sub.u  (9).

(11) Further, the step 2 includes the following specific processes: Step 2.1.1, the LS-SVM regression model is built. Specifically, a Lagrangian function is introduced for solving it. A radial basis function K(x,x.sub.i)=exp(−∥x−x.sub.i∥.sub.2/2σ.sup.2) is selected, and σ is a kernel width. The LS-SVM regression model is obtained:

(12) $\begin{array}{cc}f\left(x\right)=\underset{i=1}{\overset{l}{.Math.}}{\alpha }_{i}K\left(x,x_i\right)+b.& \left(5\right)\end{array}$ Step 2.1.2, whether a historical optimum adaptive degree P.sub.b meets a constraint condition or whether the number of iterations reaches the maximum is checked. If the constraint condition is still not met, and the number of iterations is not the maximum, a step 2.1.3 is performed, and otherwise, a result is mapped into the kernel function σ and the penalty coefficient c of the LS-SVM model. Step 2.1.3, speed and positions of particles are regulated, and an inertia weight is regulated.

(13) Further, in the step 2.1.3, a self-adaptative regulation inertia weight method is used to regulate the inertia weight:

(14) $\begin{array}{cc}w=\{\begin{array}{cc}{w}_{\min }-\frac{\left(w_{\max }-w_{\min }\right)\left(f-f_{\min }\right)}{f_{\mathrm{avg}}-f_{\min }},& f<f_{avg}\\ w_{\max },& f\ge f_{\mathrm{avg}}\end{array}.& \left(10\right)\end{array}$

(15) In the formula, w.sub.min is a minimum value of w. w.sub.max is a maximum value of w. f is an adaptive degree of a current particle. f.sub.avg is an average adaptive value of all particles. f.sub.min is a minimum adaptive value of all particles.

(16) Further, a specific process for calculating the residual r.sub.i of the current state in the step 3.1 is as follows:
r.sub.i=y.sub.i−f(x.sub.i)  (11).

(17) In the formula, y.sub.i is a true value in a sample set, and f(x.sub.i) is a predicated value of the LS-SVM regression model after optimization by the improved particle swarm algorithm.

(18) Further, a specific process for calculating the similarity trend t.sub.i of the current state in the step 3.2 is as follows:

(19) $\begin{array}{cc}{t}_i=\frac{1}{n}\sqrt{\underset{j=1}{\overset{n}{.Math.}}{\left(x_i-X_j\right)}^2}.& \left(12\right)\end{array}$

(20) In the formula, x.sub.i is a coordinate of the current state, and X.sub.j is a coordinate of the j-th clustering center.

(21) Further, a specific process for calculating the risk coefficient d.sub.i in the step 3.3 is as follows:
d.sub.i=ar.sub.i+bt.sub.i  (13).

(22) In the formula, a and b are weight factors, and are initialized to 0.5 and 0.5 according to historical data.

(23) Further, the step of translating the winner nerve cell s and all topological neighbors thereof refers to all nerve cells having connections with the winner nerve cell s.

(24) As a preference, in the step 1.5, if the two optimum nerve cells s and t are connected, the age of the connection is set to be zero, and otherwise, a connection is created between the two optimum nerve cells.

(25) As a preference, in the step 1, a growing neural gas (GNG) algorithm is used to calculate the feature vector of the historical normal operation state.

(26) The embodiment is only directed to illustrate the technical idea of the present invention, but are not to be considered to limit the protection scope of the present invention. Technical ideas provided according to the present invention, and any modification made on the basis of the technical scheme shall fall within the protection scope of the present invention.